Skip to content

GitLab

Commit cba8f7f2 authored by Anthony Chang's avatar Anthony Chang
Browse files

Merge remote-tracking branch 'upstream/develop' into gemm-layernorm-4

parents cc50b687 b653c5eb
Hide whitespace changes
Inline Side-by-side
Showing with 64 additions and 4860 deletions
library/include/ck/library/host_tensor/host_reduction.hpp
View file @ cba8f7f2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*******************************************************************************
*
* MIT License
*
* Copyright (c) 2020 Advanced Micro Devices, Inc.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
*******************************************************************************/
#ifndef HOST_REDUCTION_HPP_
#define HOST_REDUCTION_HPP_
#pragma once
#include <vector>
#include <array>
#include <functional>
#include "reduction_enums.hpp"
#include "reduction_common.hpp"
#include "host_common_util.hpp"
#include "host_tensor.hpp"
#include "data_type.hpp"
#include "reduction_functions_accumulate.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/utility/reduction_common.hpp"
#include "ck/utility/reduction_functions_accumulate.hpp"
#include "ck/library/host_tensor/host_common_util.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
template <int NDim>
static void get_all_indexes(const std::array<size_t, NDim>& dimLengths,
......@@ -400,5 +376,3 @@ struct ReductionHost
};
};
};
#endif
library/include/ck/library/host_tensor/host_tensor.hpp
View file @ cba8f7f2
#ifndef HOST_TENSOR_HPP
#define HOST_TENSOR_HPP
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <thread>
#include <vector>
......@@ -8,7 +10,8 @@
#include <utility>
#include <cassert>
#include <iostream>
#include "data_type.hpp"
#include "ck/utility/data_type.hpp"
template <typename Range>
std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim)
......@@ -107,6 +110,11 @@ struct HostTensorDescriptor
return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
}
std::size_t GetOffsetFromMultiIndex(std::vector<std::size_t> iss) const
{
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:
......@@ -223,6 +231,8 @@ struct Tensor
return ret;
}
Tensor(const Tensor& other) : mDesc(other.mDesc), mData(other.mData) {}
template <typename F>
void ForEach_impl(F&& f, std::vector<size_t>& idx, size_t rank)
{
......@@ -246,6 +256,29 @@ struct Tensor
ForEach_impl(std::forward<F>(f), idx, size_t(0));
}
template <typename F>
void ForEach_impl(const F&& f, std::vector<size_t>& idx, size_t rank) const
{
if(rank == mDesc.GetNumOfDimension())
{
f(*this, idx);
return;
}
// else
for(size_t i = 0; i < mDesc.GetLengths()[rank]; i++)
{
idx[rank] = i;
ForEach_impl(std::forward<const F>(f), idx, rank + 1);
}
}
template <typename F>
void ForEach(const F&& f) const
{
std::vector<size_t> idx(mDesc.GetNumOfDimension(), 0);
ForEach_impl(std::forward<const F>(f), idx, size_t(0));
}
template <typename G>
void GenerateTensorValue(G g, std::size_t num_thread = 1)
{
......@@ -306,6 +339,16 @@ struct Tensor
return mData[mDesc.GetOffsetFromMultiIndex(is...)];
}
T& operator()(std::vector<std::size_t> idx)
{
return mData[mDesc.GetOffsetFromMultiIndex(idx)];
}
const T& operator()(std::vector<std::size_t> idx) const
{
return mData[mDesc.GetOffsetFromMultiIndex(idx)];
}
typename std::vector<T>::iterator begin() { return mData.begin(); }
typename std::vector<T>::iterator end() { return mData.end(); }
......@@ -319,7 +362,8 @@ struct Tensor
};
template <typename X>
HostTensorDescriptor::HostTensorDescriptor(const std::vector<X>& lens) : mLens(lens)
HostTensorDescriptor::HostTensorDescriptor(const std::vector<X>& lens)
: mLens(lens.begin(), lens.end())
{
this->CalculateStrides();
}
......@@ -327,7 +371,7 @@ HostTensorDescriptor::HostTensorDescriptor(const std::vector<X>& lens) : mLens(l
template <typename X, typename Y>
HostTensorDescriptor::HostTensorDescriptor(const std::vector<X>& lens,
const std::vector<Y>& strides)
: mLens(lens), mStrides(strides)
: mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
{
}
......@@ -383,5 +427,3 @@ float check_error(const Tensor<T>& ref, const Tensor<T>& result)
return linf_error;
}
#endif
library/include/ck/library/host_tensor/host_tensor_generator.hpp
View file @ cba8f7f2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cmath>
#include <numeric>
#include "config.hpp"
#include "ck/ck.hpp"
template <typename T>
struct GeneratorTensor_0
......@@ -18,12 +21,12 @@ struct GeneratorTensor_0
template <typename T>
struct GeneratorTensor_1
{
int value = 1;
T value = 1;
template <typename... Is>
T operator()(Is...)
{
return ck::type_convert<T>(value);
return value;
}
};
......
library/include/ck/library/obselete_driver_offline/debug.hpp deleted 100644 → 0
View file @ cc50b687
#ifndef DEBUG_HPP
#define DEBUG_HPP
namespace debug {
namespace debug_driver_gemm_xdlops_v2r3 {
// these vars are on host, they control block_id to C matrix tile idx (m0, n0) mapping
static ck::index_t M01 = 1;
static ck::index_t N01 = 1;
} // namespace debug_driver_gemm_xdlops_v2r3
} // namespace debug
#endif
library/include/ck/library/obselete_driver_offline/device_convolution_add_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "driver_convolution_add_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
ck::ActivTypeEnum activ_type,
typename InLengths,
typename WeiLengths,
typename AddLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_add_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1(
const InLengths& in_n_c0_hi_wi_c1_lengths,
const WeiLengths& wei_k_c0_y_x_c1_lengths,
const AddLengths& add_n_k0_hox2_wox2_k1_lengths,
const OutLengths& out_n_k0_ho_wo_k1_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_c0_hi_wi_c1,
const Tensor<TInWei>& wei_k_c0_y_x_c1,
const Tensor<TOut>& bias_k0_k1,
const Tensor<TOut>& add_n_k0_hox2_wox2_k1,
Tensor<TOut>& add_n_k0_hox2_wox2_k1_out,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
const auto N = out_n_k0_ho_wo_k1_lengths[I0];
const auto K0 = out_n_k0_ho_wo_k1_lengths[I1];
const auto Ho = out_n_k0_ho_wo_k1_lengths[I2];
const auto Wo = out_n_k0_ho_wo_k1_lengths[I3];
const auto K1 = out_n_k0_ho_wo_k1_lengths[I4];
const auto C0 = in_n_c0_hi_wi_c1_lengths[I1];
const auto Hi = in_n_c0_hi_wi_c1_lengths[I2];
const auto Wi = in_n_c0_hi_wi_c1_lengths[I3];
const auto C1 = in_n_c0_hi_wi_c1_lengths[I4];
const auto K = wei_k_c0_y_x_c1_lengths[I0];
const auto Y = wei_k_c0_y_x_c1_lengths[I2];
const auto X = wei_k_c0_y_x_c1_lengths[I3];
const auto Hox2 = add_n_k0_hox2_wox2_k1_lengths[I2];
const auto Wox2 = add_n_k0_hox2_wox2_k1_lengths[I3];
DeviceMem in_n_c0_hi_wi_c1_device_buf(sizeof(TInWei) *
in_n_c0_hi_wi_c1.mDesc.GetElementSpace());
DeviceMem wei_k_c0_y_x_c1_device_buf(sizeof(TInWei) * wei_k_c0_y_x_c1.mDesc.GetElementSpace());
DeviceMem bias_k0_k1_device_buf(sizeof(TOut) * bias_k0_k1.mDesc.GetElementSpace());
DeviceMem add_n_k0_hox2_wox2_k1_device_buf(sizeof(TOut) *
add_n_k0_hox2_wox2_k1.mDesc.GetElementSpace());
in_n_c0_hi_wi_c1_device_buf.ToDevice(in_n_c0_hi_wi_c1.mData.data());
wei_k_c0_y_x_c1_device_buf.ToDevice(wei_k_c0_y_x_c1.mData.data());
bias_k0_k1_device_buf.ToDevice(bias_k0_k1.mData.data());
add_n_k0_hox2_wox2_k1_device_buf.ToDevice(add_n_k0_hox2_wox2_k1.mData.data());
constexpr index_t InWeiVectorSize = 8;
if(C1 % InWeiVectorSize != 0)
{
throw std::runtime_error("wrong! C1 cannot be divided by InWeiVectorSize");
}
#if 0
constexpr index_t BlockSize = 256;
constexpr index_t KPerBlock = 32;
constexpr index_t HoPerBlock = 8;
constexpr index_t WoPerBlock = 64;
constexpr index_t E1 = C0 * 9;
constexpr index_t E2 = 1;
constexpr index_t E1PerBlock = C0;
constexpr index_t KPerThread = 16;
constexpr index_t HoPerThread = 2;
constexpr index_t WoPerThread = 2;
constexpr index_t EPerThread = 1;
using ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2 = Sequence<1, 9, 1, E2>;
using ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2 = Sequence<1, E1PerBlock, KPerBlock, 1>;
constexpr index_t ABlockTransferSrcScalarPerVector_E2 = E2;
constexpr index_t ABlockTransferDstScalarPerVector_E2 = E2;
constexpr index_t BThreadTransferSrcScalarPerVector_E2 = E2;
constexpr index_t CThreadTransferDstScalarPerVector_K = K1;
#elif 1
constexpr auto BlockSize = 64;
constexpr auto KPerBlock = 8;
constexpr auto HoPerBlock = 8;
constexpr auto WoPerBlock = 32;
constexpr auto E1 = 2 * 9;
constexpr auto E2 = 1;
constexpr auto K2 = 2;
constexpr auto E1PerBlock = 2;
constexpr auto KPerThread = KPerBlock;
constexpr auto HoPerThread = 2;
constexpr auto WoPerThread = 2;
constexpr auto EPerThread = 1;
using ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2 = Sequence<1, 9, 1, 1, E2>;
using ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2 =
Sequence<1, E1PerBlock, 1, KPerBlock, 1>;
constexpr auto ABlockTransferSrcScalarPerVector_E2 = E2;
constexpr auto ABlockTransferDstScalarPerVector_E2 = E2;
constexpr auto BThreadTransferSrcScalarPerVector_E2 = E2;
constexpr auto CThreadTransferDstScalarPerVector_K = InWeiVectorSize;
#endif
const auto in_n_c0_hi_wi_c1_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, C0, Hi, Wi, E2));
const auto wei_k_c0_y_x_c1_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, C0, Y, X, E2));
const auto add_n_k0_hox2_wox2_k1_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, K0, Hox2, Wox2, K1));
const auto out_n_k0_ho_wo_k1_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, K0, Ho, Wo, K1));
constexpr auto conv_driver =
DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nc0hwc1_kc0yxc1_nk0hwk1_add<
BlockSize,
typename vector_type<TInWei, InWeiVectorSize>::type,
TAcc,
TOut,
E1,
E2,
K2,
KPerBlock,
HoPerBlock,
WoPerBlock,
E1PerBlock,
KPerThread,
HoPerThread,
WoPerThread,
EPerThread,
ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2,
ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2,
ABlockTransferSrcScalarPerVector_E2,
ABlockTransferDstScalarPerVector_E2,
BThreadTransferSrcScalarPerVector_E2,
CThreadTransferDstScalarPerVector_K,
activ_type>{};
std::cerr << "conv_bias_activ_resize_add_input_"
<< "n" << N << "c" << C0 << "h" << Hi << "w" << Wi << "c" << C1 << "_filter_k" << K
<< "c" << C0 << "y" << Y << "x" << X << "c" << C1 << "_addout_n" << N << "k" << K0
<< "h" << Ho * 2 << "w" << Wo * 2 << "k" << K1 << std::endl;
for(int i = 0; i < 5; i++)
{
const auto ave_time =
conv_driver.Run(wei_k_c0_y_x_c1_desc,
in_n_c0_hi_wi_c1_desc,
out_n_k0_ho_wo_k1_desc,
add_n_k0_hox2_wox2_k1_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
wei_k_c0_y_x_c1_device_buf.GetDeviceBuffer()),
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
in_n_c0_hi_wi_c1_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(bias_k0_k1_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(add_n_k0_hox2_wox2_k1_device_buf.GetDeviceBuffer()),
nrepeat);
{
float perf = static_cast<float>(std::size_t(2) * N * K * Ho * Wo * C0 * C1 * Y * X) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
add_n_k0_hox2_wox2_k1_device_buf.ToDevice(add_n_k0_hox2_wox2_k1.mData.data());
conv_driver.Run(wei_k_c0_y_x_c1_desc,
in_n_c0_hi_wi_c1_desc,
out_n_k0_ho_wo_k1_desc,
add_n_k0_hox2_wox2_k1_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
wei_k_c0_y_x_c1_device_buf.GetDeviceBuffer()),
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
in_n_c0_hi_wi_c1_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(bias_k0_k1_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(add_n_k0_hox2_wox2_k1_device_buf.GetDeviceBuffer()),
0);
add_n_k0_hox2_wox2_k1_device_buf.FromDevice(add_n_k0_hox2_wox2_k1_out.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_backward_data_convolution_into_gemm_v4r1_nhwc_kyxc_nhwk.hpp"
#include "driver_gemm_xdlops_v2r3.hpp"
#include "debug.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
const InLengths& in_n_hi_wi_c_lengths,
const WeiLengths& wei_k_y_x_c_lengths,
const OutLengths& out_n_ho_wo_k_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
Tensor<TInWei>& in_n_hi_wi_c,
const Tensor<TInWei>& wei_k_y_x_c,
const Tensor<TOut>& out_n_ho_wo_k,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
DeviceMem in_n_hi_wi_c_device_buf(sizeof(TInWei) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_k_y_x_c_device_buf(sizeof(TInWei) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
#if 0
// [M, N, K0, K1] = [128, 128, 4, 4], C = 64, for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 2;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#elif 0
// [M, N, K0, K1] = [128, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 2;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#elif 1
// [M, N, K0, K1] = [256, 128, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#elif 1
// [M, N, K0, K1] = [128, 256, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 4;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 2;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#endif
const auto descs =
transform_backward_data_convolution_into_gemm_v4r1_nhwc_kyxc_nhwk(wei_k_y_x_c_desc,
out_n_ho_wo_k_desc,
in_n_hi_wi_c_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
I0,
I0,
Number<GemmK1>{});
const auto wei_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto out_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto in_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto out_gemmk0_gemmn_gemmk1_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0>{}, // 1+: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0>{}, // 1-: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 2-: GemmK1
// clang-format off
constexpr auto in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks = make_tuple(
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
//clang-format on
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{};
constexpr auto out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0>{};
for(index_t i = 0; i < 5; ++i)
{
float ave_time = driver_gemm_xdlops_v2r3<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum::Set,
decltype(wei_gemmk0_gemmm_gemmk1_grid_desc),
decltype(out_gemmk0_gemmn_gemmk1_grid_desc),
decltype(in_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerXDL,
GemmNPerXDL,
GemmK1,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<2, 0, 1>,
Sequence<0, 2, 1>,
1,
GemmABlockTransferSrcScalarPerVector_GemmM,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmBBlockTransferSrcScalarPerVector_GemmK1,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
Sequence<1, 3, 7, 0, 2, 4, 5, 6>,
6,
GemmCThreadTransferDstScalarPerVector,
decltype(wei_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(out_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false, // CAccessOrderMRepeatNRepeat
false, // ABlockLdsExtraM
false // BBlockLdsExtraN
>(static_cast<TInWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
wei_gemmk0_gemmm_gemmk1_grid_desc,
out_gemmk0_gemmn_gemmk1_grid_desc,
in_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
wei_gemmk0_gemmm_gemmk1_grid_step_hacks,
out_gemmk0_gemmn_gemmk1_grid_step_hacks,
in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
{
const auto N = out_n_ho_wo_k_lengths[I0];
const auto K = out_n_ho_wo_k_lengths[I3];
const auto C = wei_k_y_x_c_lengths[I3];
const auto Ho = out_n_ho_wo_k_lengths[I1];
const auto Wo = out_n_ho_wo_k_lengths[I2];
const auto Y = wei_k_y_x_c_lengths[I1];
const auto X = wei_k_y_x_c_lengths[I2];
float perf = static_cast<float>((std::size_t(2) * N * K * Ho * Wo * C * Y * X)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
// copy result back to host
in_n_hi_wi_c_device_buf.FromDevice(in_n_hi_wi_c.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk.hpp"
#include "driver_gemm_xdlops_v2r3.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk(
const InLengths& in_n_hi_wi_c_lengths,
const WeiLengths& wei_k_y_x_c_lengths,
const OutLengths& out_n_ho_wo_k_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
Tensor<TInWei>& in_n_hi_wi_c,
const Tensor<TInWei>& wei_k_y_x_c,
const Tensor<TOut>& out_n_ho_wo_k,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
DeviceMem in_n_hi_wi_c_device_buf(sizeof(TInWei) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_k_y_x_c_device_buf(sizeof(TInWei) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
#if 0
// [M, N, K0, K1] = [256, 128, 4, 4], C = 128, for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 128, 4, 4], C = 64, for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [256, 128, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [128, 256, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 4;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [128, 128, 4, 8], C = 64, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 64, 4, 8], C = 64, for fp16
constexpr index_t BlockSize = 128;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 64;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 32, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 64, 4, 8], C = 32, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 64;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto out_gemmk0_gemmm_gemmk1_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 0+: gemmk0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0>{}, // 1+: gemmm
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 2+: gemmk1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 0-: gemmk0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0>{}, // 1-: gemmm
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 2-:
// gemmk1
constexpr auto wei_gemmk0_gemmn_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 0+: gemmk0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: gemmn
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 2+: gemmk1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 0-: Gemmk0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: Gemmn
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 2-: Gemmk1
// clang-format off
constexpr auto in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks = make_tuple(
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
// clang-format on
constexpr auto out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0>{};
constexpr auto wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{};
for(index_t i = 0; i < 5; ++i)
{
const auto ConvStrideH = conv_strides[I0];
const auto ConvStrideW = conv_strides[I1];
const auto ConvDilationH = conv_dilations[I0];
const auto ConvDilationW = conv_dilations[I1];
const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
const auto YTilde = ConvStrideH / GcdStrideDilationH;
const auto XTilde = ConvStrideW / GcdStrideDilationW;
float ave_time = 0;
for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
{
for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
{
const auto descs =
transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk(
out_n_ho_wo_k_desc,
wei_k_y_x_c_desc,
in_n_hi_wi_c_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
i_ytilde,
i_xtilde,
Number<GemmK1>{});
const auto out_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto wei_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto in_gemmm_gemmn_grid_desc = descs[I2];
const auto GemmK0 = out_gemmk0_gemmm_gemmk1_grid_desc.GetLength(I0);
if(GemmK0 != 0)
{
ave_time += driver_gemm_xdlops_v2r3<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum::Set,
decltype(out_gemmk0_gemmm_gemmk1_grid_desc),
decltype(wei_gemmk0_gemmn_gemmk1_grid_desc),
decltype(in_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerWave,
GemmNPerWave,
GemmK1,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmABlockTransferSrcScalarPerVector_GemmK1,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<2, 0, 1>,
Sequence<0, 2, 1>,
1,
GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
#if 0
Sequence<0, 2, 4, 5, 6, 1, 3, 7>,
#else
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
#endif
7,
GemmCThreadTransferDstScalarPerVector,
decltype(out_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
true, // CAccessOrderMRepeatNRepeat
false, // ABlockLdsExtraM
false // BBlockLdsExtraN
>(static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
out_gemmk0_gemmm_gemmk1_grid_desc,
wei_gemmk0_gemmn_gemmk1_grid_desc,
in_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
out_gemmk0_gemmm_gemmk1_grid_step_hacks,
wei_gemmk0_gemmn_gemmk1_grid_step_hacks,
in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
}
}
}
{
const auto N = out_n_ho_wo_k_lengths[I0];
const auto K = out_n_ho_wo_k_lengths[I3];
const auto C = wei_k_y_x_c_lengths[I3];
const auto Ho = out_n_ho_wo_k_lengths[I1];
const auto Wo = out_n_ho_wo_k_lengths[I2];
const auto Y = wei_k_y_x_c_lengths[I1];
const auto X = wei_k_y_x_c_lengths[I2];
float perf = static_cast<float>((std::size_t(2) * N * K * Ho * Wo * C * Y * X)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
// copy result back to host
in_n_hi_wi_c_device_buf.FromDevice(in_n_hi_wi_c.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk_1x1.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk.hpp"
#include "driver_gemm_xdlops_v2r3.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk_1x1(
const InLengths& in_n_hi_wi_c_lengths,
const WeiLengths& wei_k_y_x_c_lengths,
const OutLengths& out_n_ho_wo_k_lengths,
const ConvStrides& conv_strides,
const ConvDilations&,
const InLeftPads&,
const InRightPads&,
Tensor<TInWei>& in_n_hi_wi_c,
const Tensor<TInWei>& wei_k_y_x_c,
const Tensor<TOut>& out_n_ho_wo_k,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
DeviceMem in_n_hi_wi_c_device_buf(sizeof(TInWei) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_k_y_x_c_device_buf(sizeof(TInWei) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
#if 0
// [M, N, K0, K1] = [256, 128, 4, 4], C = 128, for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 128, 4, 4], C = 64, for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [256, 128, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [128, 256, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 4;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 128, 4, 8], C = 64, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 64, 4, 8], C = 64, for fp16
constexpr index_t BlockSize = 128;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 64;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 32, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 64, 4, 8], C = 32, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 64;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto out_gemmk0_gemmm_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0>{}, // 0+: gemmk0
Sequence<0, 0, 0>{}, // 1+: gemmm
Sequence<0, 0, 0>{}), // 2+: gemmk1
make_tuple(Sequence<0, 0, 0>{}, // 0-: gemmk0
Sequence<0, 0, 0>{}, // 1-: gemmm
Sequence<0, 0, 0>{})); // 2-: gemmk1
constexpr auto wei_gemmk0_gemmn_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0>{}, // 0+: gemmk0
Sequence<0, 0, 0>{}, // 1+: gemmn
Sequence<0, 0, 0>{}), // 2+: gemmk1
make_tuple(Sequence<0, 0, 0>{}, // 0-: Gemmk0
Sequence<0, 0, 0>{}, // 1-: Gemmn
Sequence<0, 0, 0>{})); // 2-: Gemmk1
// clang-format off
constexpr auto in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks = make_tuple(
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
// clang-format on
constexpr auto out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0>{};
constexpr auto wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0>{};
for(index_t i = 0; i < 5; ++i)
{
const auto descs = transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk_1x1(
out_n_ho_wo_k_desc,
wei_k_y_x_c_desc,
in_n_hi_wi_c_desc,
conv_strides,
Number<GemmK1>{});
const auto out_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto wei_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto in_gemmm_gemmn_grid_desc = descs[I2];
float ave_time = driver_gemm_xdlops_v2r3<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum::Set,
decltype(out_gemmk0_gemmm_gemmk1_grid_desc),
decltype(wei_gemmk0_gemmn_gemmk1_grid_desc),
decltype(in_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerWave,
GemmNPerWave,
GemmK1,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmABlockTransferSrcScalarPerVector_GemmK1,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<2, 0, 1>,
Sequence<0, 2, 1>,
1,
GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
#if 0
Sequence<0, 2, 4, 5, 6, 1, 3, 7>,
#else
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
#endif
7,
GemmCThreadTransferDstScalarPerVector,
decltype(out_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
true, // CAccessOrderMRepeatNRepeat
false, // ABlockLdsExtraM
false // BBlockLdsExtraN
>(static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
out_gemmk0_gemmm_gemmk1_grid_desc,
wei_gemmk0_gemmn_gemmk1_grid_desc,
in_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
out_gemmk0_gemmm_gemmk1_grid_step_hacks,
wei_gemmk0_gemmn_gemmk1_grid_step_hacks,
in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
{
const auto N = out_n_ho_wo_k_lengths[I0];
const auto K = out_n_ho_wo_k_lengths[I3];
const auto C = wei_k_y_x_c_lengths[I3];
const auto Ho = out_n_ho_wo_k_lengths[I1];
const auto Wo = out_n_ho_wo_k_lengths[I2];
const auto Y = wei_k_y_x_c_lengths[I1];
const auto X = wei_k_y_x_c_lengths[I2];
float perf = static_cast<float>((std::size_t(2) * N * K * Ho * Wo * C * Y * X)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
// copy result back to host
in_n_hi_wi_c_device_buf.FromDevice(in_n_hi_wi_c.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_atomic_nchw_kcyx_nkhw.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp"
#include "driver_gemm_xdlops_v2r4.hpp"
template <typename TIn,
typename TWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads,
typename GridSizeType>
void device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_atomic_nchw_kcyx_nkhw(
const InLengths& in_n_c_hi_wi_lengths,
const WeiLengths& wei_k_c_y_x_lengths,
const OutLengths& out_n_k_ho_wo_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TIn>& in_n_c_hi_wi,
Tensor<TWei>& wei_k_c_y_x,
const Tensor<TOut>& out_n_k_ho_wo,
GridSizeType desired_grid_size,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
DeviceMem in_n_c_hi_wi_device_buf(sizeof(TIn) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_k_c_y_x_device_buf(sizeof(TWei) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_n_k_ho_wo_device_buf(sizeof(TOut) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_n_c_hi_wi_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
wei_k_c_y_x_device_buf.ToDevice(wei_k_c_y_x.mData.data());
out_n_k_ho_wo_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
const auto in_n_c_hi_wi_desc = make_naive_tensor_descriptor_packed(in_n_c_hi_wi_lengths);
const auto wei_k_c_y_x_desc = make_naive_tensor_descriptor_packed(wei_k_c_y_x_lengths);
const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths);
#if 1
// [M, N, K0, K1] = [128, 128, 4, 8] for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmB_GemmK0_GemmM_GemmK1 = Sequence<1, 1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmB_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 64, 1>;
// using vector load 4, so config's wo*ho must be a multiple of 4
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmB_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmB_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif
const auto N = in_n_c_hi_wi_desc.GetLength(I0);
const auto C = in_n_c_hi_wi_desc.GetLength(I1);
const auto K = out_n_k_ho_wo_desc.GetLength(I1);
const auto Ho = out_n_k_ho_wo_desc.GetLength(I2);
const auto Wo = out_n_k_ho_wo_desc.GetLength(I3);
const auto Y = wei_k_c_y_x_desc.GetLength(I2);
const auto X = wei_k_c_y_x_desc.GetLength(I3);
const auto GemmM = K;
const auto GemmN = Y * X * C;
const auto GemmKTotal = N * Ho * Wo;
const auto GridMN = GemmM * GemmN / (GemmMPerBlock * GemmNPerBlock);
const index_t GemmKBatch = std::max(desired_grid_size / GridMN, 1);
const index_t GemmK0 =
math::integer_divide_ceil(GemmKTotal, GemmK1 * GemmKPerBlock * GemmKBatch) * GemmKPerBlock;
const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1;
std::cout << "GemmKTotal: " << GemmKTotal << " GrideSizeMN: " << GridMN
<< " GemmKBatch: " << GemmKBatch << " GemmK0: " << GemmK0 << " gemmKPad: " << GemmKPad
<< std::endl;
const auto descs =
transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw_pad(
wei_k_c_y_x_desc,
in_n_c_hi_wi_desc,
out_n_k_ho_wo_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<GemmK1>{},
GemmKBatch,
GemmKPad);
const auto out_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto in_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto wei_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto out_gemmk0_gemmm_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 1, 0, 0, 0, 0>{}, // 0+: GemmB
Sequence<0, 0, 1, 0, 0, 0, 0>{}, // 1+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 2+: GemmM
Sequence<0, 0, 1, 0, 0, 0, 0>{}), // 3+: GemmK1
make_tuple(Sequence<0, 0, 2, 0, 0, 0, 0>{}, // 0-: GemB
Sequence<0, 0, 2, 0, 0, 0, 0>{}, // 1-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 2-: GemmM
Sequence<0, 0, 2, 0, 0, 0, 0>{})); // 3-: GemmK1
constexpr auto in_gemmk0_gemmn_gemmk1_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}, // 0+: GemmB
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}, // 1+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 2+: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}), // 3+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{}, // 0-: GemmB
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{}, // 1-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 2-: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{})); // 3-: GemmK1
constexpr auto wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
constexpr auto out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 1, 0, 0, 0, 0>{};
constexpr auto in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 0, 0, 0, 0>{};
const auto driver_gemm_xdlops =
driver_gemm_xdlops_v2r4<BlockSize,
TIn,
TAcc,
TWei,
InMemoryDataOperationEnum::AtomicAdd,
decltype(out_gemmk0_gemmm_gemmk1_grid_desc),
decltype(in_gemmk0_gemmn_gemmk1_grid_desc),
decltype(wei_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerWave,
GemmNPerWave,
GemmK1,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmB_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmB_GemmK0_GemmM_GemmK1,
Sequence<0, 2, 1, 3>,
Sequence<0, 2, 1, 3>,
3,
GemmABlockTransferSrcScalarPerVector_GemmK1,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmB_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmB_GemmK0_GemmN_GemmK1,
Sequence<0, 2, 1, 3>,
Sequence<0, 2, 1, 3>,
3,
GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
Sequence<3, 0, 1, 2, 7, 5, 4, 6>,
7,
GemmCThreadTransferDstScalarPerVector,
decltype(out_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false,
true,
true>;
for(index_t i = 0; i < 5; ++i)
{
float ave_time =
driver_gemm_xdlops(static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
static_cast<TIn*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
static_cast<TWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
out_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
out_gemmk0_gemmm_gemmk1_grid_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_step_hacks,
wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
float perf = static_cast<float>(calculate_convolution_flops(
in_n_c_hi_wi_desc, wei_k_c_y_x_desc, out_n_k_ho_wo_desc)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
}
wei_k_c_y_x_device_buf.ToDevice(wei_k_c_y_x.mData.data());
driver_gemm_xdlops(static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
static_cast<TIn*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
static_cast<TWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
out_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
out_gemmk0_gemmm_gemmk1_grid_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_step_hacks,
wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
0);
// copy result back to host
wei_k_c_y_x_device_buf.FromDevice(wei_k_c_y_x.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_backward_weight_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp"
#include "driver_gemm_xdlops_v2r3.hpp"
template <typename TIn,
typename TWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw(
const InLengths& in_n_c_hi_wi_lengths,
const WeiLengths& wei_k_c_y_x_lengths,
const OutLengths& out_n_k_ho_wo_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TIn>& in_n_c_hi_wi,
Tensor<TWei>& wei_k_c_y_x,
const Tensor<TOut>& out_n_k_ho_wo,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
DeviceMem in_n_c_hi_wi_device_buf(sizeof(TIn) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_k_c_y_x_device_buf(sizeof(TWei) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_n_k_ho_wo_device_buf(sizeof(TOut) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_n_c_hi_wi_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
wei_k_c_y_x_device_buf.ToDevice(wei_k_c_y_x.mData.data());
out_n_k_ho_wo_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
const auto in_n_c_hi_wi_desc = make_naive_tensor_descriptor_packed(in_n_c_hi_wi_lengths);
const auto wei_k_c_y_x_desc = make_naive_tensor_descriptor_packed(wei_k_c_y_x_lengths);
const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths);
#if 0
// [M, N, K0, K1] = [128, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
// using vector load 4, so config's wo*ho must be a multiple of 4
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [128, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
// using vector load 4, so config's wo*ho must be a multiple of 4
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif
const auto descs = transform_backward_weight_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw_pad(
wei_k_c_y_x_desc,
in_n_c_hi_wi_desc,
out_n_k_ho_wo_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<GemmK1>{});
const auto out_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto in_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto wei_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto out_gemmk0_gemmm_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 1, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1+: GemmM
Sequence<0, 0, 1, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 2, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1-: GemmM
Sequence<0, 0, 2, 0, 0>{})); // 2-: GemmK1
constexpr auto in_gemmk0_gemmn_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, // 1+: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, // 1-: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{})); // 2-: GemmK1
constexpr auto wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
constexpr auto out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 1, 0, 0>{};
constexpr auto in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 0, 0>{};
for(index_t i = 0; i < 5; ++i)
{
float ave_time = driver_gemm_xdlops_v2r3<
BlockSize,
TIn,
TAcc,
TWei,
InMemoryDataOperationEnum::Set,
decltype(out_gemmk0_gemmm_gemmk1_grid_desc),
decltype(in_gemmk0_gemmn_gemmk1_grid_desc),
decltype(wei_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerWave,
GemmNPerWave,
GemmK1,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmABlockTransferSrcScalarPerVector_GemmK1,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
Sequence<3, 0, 1, 2, 7, 5, 4, 6>,
7,
GemmCThreadTransferDstScalarPerVector,
decltype(out_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false, // CAccessOrderMRepeatNRepeat
true, // ABlockLdsExtraM
true // BBlockLdsExtraN
>(static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
static_cast<TIn*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
static_cast<TWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
out_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
out_gemmk0_gemmm_gemmk1_grid_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_step_hacks,
wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
float perf = static_cast<float>(calculate_convolution_flops(
in_n_c_hi_wi_desc, wei_k_c_y_x_desc, out_n_k_ho_wo_desc)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
}
// copy result back to host
wei_k_c_y_x_device_buf.FromDevice(wei_k_c_y_x.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_atomic_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk.hpp"
#include "driver_gemm_xdlops_v2r4.hpp"
template <typename TIn,
typename TWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads,
typename GridSizeType>
void device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_atomic_nhwc_kyxc_nhwk(
const InLengths& in_n_hi_wi_c_lengths,
const WeiLengths& wei_k_y_x_c_lengths,
const OutLengths& out_n_ho_wo_k_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TIn>& in_n_hi_wi_c,
Tensor<TWei>& wei_k_y_x_c,
const Tensor<TOut>& out_n_ho_wo_k,
GridSizeType desired_grid_size,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
DeviceMem in_n_hi_wi_c_device_buf(sizeof(TIn) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_k_y_x_c_device_buf(sizeof(TWei) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
#if 0
// [M, N, K0, K1] = [128, 256, 4, 4] for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 4;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 1, 4, 2>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 32, 2>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 8, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#elif 1
// [M, N, K0, K1] = [128, 128, 4, 4] for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 1, 4, 2>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 32, 2>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 4, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif
const auto N = in_n_hi_wi_c_desc.GetLength(I0);
const auto C = in_n_hi_wi_c_desc.GetLength(I3);
const auto K = out_n_ho_wo_k_desc.GetLength(I3);
const auto Ho = out_n_ho_wo_k_desc.GetLength(I1);
const auto Wo = out_n_ho_wo_k_desc.GetLength(I2);
const auto Y = wei_k_y_x_c_desc.GetLength(I1);
const auto X = wei_k_y_x_c_desc.GetLength(I2);
const auto GemmM = Y * X * C;
const auto GemmN = K;
const auto GemmKTotal = N * Ho * Wo;
const auto GridMN = GemmM * GemmN / (GemmMPerBlock * GemmNPerBlock);
const index_t GemmKBatch = std::max(desired_grid_size / GridMN, 1);
const index_t GemmK0 =
math::integer_divide_ceil(GemmKTotal, GemmK1 * GemmKPerBlock * GemmKBatch) * GemmKPerBlock;
const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1;
std::cout << "GemmKTotal: " << GemmKTotal << " GrideSizeMN: " << GridMN
<< " GemmKBatch: " << GemmKBatch << " GemmK0: " << GemmK0 << " gemmKPad: " << GemmKPad
<< std::endl;
const auto descs =
transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk_pad(
in_n_hi_wi_c_desc,
wei_k_y_x_c_desc,
out_n_ho_wo_k_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<GemmK1>{},
GemmKBatch,
GemmKPad);
const auto in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto wei_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}, // 0+: GemmKBatch
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}, // 1+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 2+: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}), // 3+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{}, // 0-: GemmKBatch
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{}, // 1-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 2-: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{})); // 3-: GemmK1
constexpr auto out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1+: GemmN
Sequence<0, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1-: GemmN
Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
constexpr auto in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 0, 0, 0, 0>{};
constexpr auto out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0>{};
const auto driver_gemm_xdlops = driver_gemm_xdlops_v2r4<
BlockSize,
TIn,
TAcc,
TWei,
InMemoryDataOperationEnum::AtomicAdd,
decltype(in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc),
decltype(out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc),
decltype(wei_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerXDL,
GemmNPerXDL,
GemmK1,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<0, 1, 2, 3>,
Sequence<0, 1, 2, 3>,
2,
GemmABlockTransferSrcScalarPerVector_GemmM,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<0, 1, 2, 3>,
Sequence<0, 1, 2, 3>,
2,
GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
6,
GemmCThreadTransferDstScalarPerVector,
decltype(in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false, // CAccessOrderMRepeatNRepeat
true,
true>;
for(index_t i = 0; i < 5; ++i)
{
float ave_time =
driver_gemm_xdlops(static_cast<TIn*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
static_cast<TWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks,
out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks,
wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
{
float perf = static_cast<float>((std::size_t(2) * N * K * Ho * Wo * C * Y * X)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
driver_gemm_xdlops(static_cast<TIn*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
static_cast<TWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks,
out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks,
wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
0);
// copy result back to host
wei_k_y_x_c_device_buf.FromDevice(wei_k_y_x_c.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp"
#include "driver_gemm_xdlops_v2r3.hpp"
#include "debug.hpp"
template <typename TIn,
typename TWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
const InLengths& in_n_hi_wi_c_lengths,
const WeiLengths& wei_k_y_x_c_lengths,
const OutLengths& out_n_ho_wo_k_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TIn>& in_n_hi_wi_c,
Tensor<TWei>& wei_k_y_x_c,
const Tensor<TOut>& out_n_ho_wo_k,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
DeviceMem in_n_hi_wi_c_device_buf(sizeof(TIn) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_k_y_x_c_device_buf(sizeof(TWei) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
#if 0
// [M, N, K0, K1] = [256, 128, 4, 4] for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 2;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [128, 128, 4, 4] for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 2>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 32, 2>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 32, 2>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif
const auto descs = transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk_pad(
in_n_hi_wi_c_desc,
wei_k_y_x_c_desc,
out_n_ho_wo_k_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<GemmK1>{});
const auto in_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto out_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto wei_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto in_gemmk0_gemmm_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, // 1+: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, // 1-: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{})); // 2-: GemmK1
constexpr auto out_gemmk0_gemmn_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1+: GemmN
Sequence<0, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1-: GemmN
Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
constexpr auto in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 0, 0>{};
constexpr auto out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0>{};
for(index_t i = 0; i < 5; ++i)
{
float ave_time = driver_gemm_xdlops_v2r3<
BlockSize,
TIn,
TAcc,
TWei,
InMemoryDataOperationEnum::Set,
decltype(in_gemmk0_gemmm_gemmk1_grid_desc),
decltype(out_gemmk0_gemmn_gemmk1_grid_desc),
decltype(wei_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerXDL,
GemmNPerXDL,
GemmK1,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<0, 2, 1>,
Sequence<0, 2, 1>,
1,
GemmABlockTransferSrcScalarPerVector_GemmM,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<0, 2, 1>,
Sequence<0, 2, 1>,
1,
GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
7,
GemmCThreadTransferDstScalarPerVector,
decltype(in_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(out_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false, // CAccessOrderMRepeatNRepeat
true,
true>(static_cast<TIn*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
static_cast<TWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
in_gemmk0_gemmm_gemmk1_grid_desc,
out_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
in_gemmk0_gemmm_gemmk1_grid_step_hacks,
out_gemmk0_gemmn_gemmk1_grid_step_hacks,
wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
{
const auto N = out_n_ho_wo_k_lengths[I0];
const auto K = out_n_ho_wo_k_lengths[I3];
const auto C = wei_k_y_x_c_lengths[I3];
const auto Ho = out_n_ho_wo_k_lengths[I1];
const auto Wo = out_n_ho_wo_k_lengths[I2];
const auto Y = wei_k_y_x_c_lengths[I1];
const auto X = wei_k_y_x_c_lengths[I2];
float perf = static_cast<float>((std::size_t(2) * N * K * Ho * Wo * C * Y * X)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
// copy result back to host
wei_k_y_x_c_device_buf.FromDevice(wei_k_y_x_c.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_backward_weight_implicit_gemm_v4r4r5_xdlops_atomic_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk.hpp"
#include "driver_gemm_xdlops_v2r4.hpp"
template <typename TIn,
typename TWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads,
typename GridSizeType>
void device_convolution_backward_weight_implicit_gemm_v4r4r5_xdlops_atomic_nhwc_kyxc_nhwk(
const InLengths& in_n_hi_wi_c_lengths,
const WeiLengths& wei_k_y_x_c_lengths,
const OutLengths& out_n_ho_wo_k_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TIn>& in_n_hi_wi_c,
Tensor<TWei>& wei_k_y_x_c,
const Tensor<TOut>& out_n_ho_wo_k,
GridSizeType desired_grid_size,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
DeviceMem in_n_hi_wi_c_device_buf(sizeof(TIn) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_k_y_x_c_device_buf(sizeof(TWei) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
#if 0
// [M, N, K0, K1] = [256, 128, 4, 4], C 128, for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 1, 8, 2>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 32, 2>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 4, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 256, 4, 4], C 128, for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 4;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 1, 4, 2>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 32, 2>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 8, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [128, 128, 4, 4], C 64, for fp32 and fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 1, 4, 2>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 32, 2>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 4, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [256, 128, 4, 8], C 128, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 1, 16, 2>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 16, 4>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 8, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 16, 4>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [128, 128, 4, 8], C 64, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 1, 8, 2>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 16, 4>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 8, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 16, 4>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 64, 4, 8], C 64, for fp16
constexpr index_t BlockSize = 128;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 64;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 1, 16, 2>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8, 4>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 8, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8, 4>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [64, 128, 4, 8], C 64, for fp16
constexpr index_t BlockSize = 128;
constexpr index_t GemmMPerBlock = 64;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 1, 8, 2>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8, 4>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 16, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8, 4>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [64, 64, 4, 8], C 32, for fp16
constexpr index_t BlockSize = 128;
constexpr index_t GemmMPerBlock = 64;
constexpr index_t GemmNPerBlock = 64;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 1, 8, 2>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8, 4>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 8, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8, 4>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif
const auto N = in_n_hi_wi_c_desc.GetLength(I0);
const auto C = in_n_hi_wi_c_desc.GetLength(I3);
const auto K = out_n_ho_wo_k_desc.GetLength(I3);
const auto Ho = out_n_ho_wo_k_desc.GetLength(I1);
const auto Wo = out_n_ho_wo_k_desc.GetLength(I2);
const auto Y = wei_k_y_x_c_desc.GetLength(I1);
const auto X = wei_k_y_x_c_desc.GetLength(I2);
const auto GemmM = K;
const auto GemmN = Y * X * C;
const auto GemmKTotal = N * Ho * Wo;
const auto GridMN = GemmM * GemmN / (GemmMPerBlock * GemmNPerBlock);
const index_t GemmKBatch = std::max(desired_grid_size / GridMN, 1);
const index_t GemmK0 =
math::integer_divide_ceil(GemmKTotal, GemmK1 * GemmKPerBlock * GemmKBatch) * GemmKPerBlock;
const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1;
std::cout << "GemmKTotal: " << GemmKTotal << " GrideSizeMN: " << GridMN
<< " GemmKBatch: " << GemmKBatch << " GemmK0: " << GemmK0 << " gemmKPad: " << GemmKPad
<< std::endl;
const auto descs = transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk_pad(
in_n_hi_wi_c_desc,
wei_k_y_x_c_desc,
out_n_ho_wo_k_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<GemmK1>{},
GemmKBatch,
GemmKPad);
const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto wei_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1+: GemmN
Sequence<0, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1-: GemmN
Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 1+: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 1-: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
constexpr auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0>{};
constexpr auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 0, 0, 0, 0>{};
const auto driver_gemm_xdlops = driver_gemm_xdlops_v2r4<
BlockSize,
TIn,
TAcc,
TWei,
InMemoryDataOperationEnum::AtomicAdd,
decltype(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc),
decltype(in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc),
decltype(wei_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerXDL,
GemmNPerXDL,
GemmK1,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<0, 1, 2, 3>,
Sequence<0, 1, 2, 3>,
2,
GemmABlockTransferSrcScalarPerVector_GemmM,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<0, 1, 2, 3>,
Sequence<0, 1, 3, 2>,
2,
GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
7,
GemmCThreadTransferDstScalarPerVector,
decltype(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false, // CAccessOrderMRepeatNRepeat
true,
true>;
// timing
for(index_t i = 0; i < 5; ++i)
{
float ave_time =
driver_gemm_xdlops(static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
static_cast<TIn*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks,
wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
{
float perf = static_cast<float>((std::size_t(2) * N * K * Ho * Wo * C * Y * X)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
// verification
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
driver_gemm_xdlops(static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
static_cast<TIn*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks,
wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
0);
// copy result back to host
wei_k_y_x_c_device_buf.FromDevice(wei_k_y_x_c.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_forward_implicit_gemm_v4r4_dlops_nchw_kcyx_nkhw.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw.hpp"
#include "driver_gemm_dlops_v1r2.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_forward_implicit_gemm_v4r4_dlops_nchw_kcyx_nkhw(
const InLengths& in_n_c_hi_wi_lengths,
const WeiLengths& wei_k_c_y_x_lengths,
const OutLengths& out_n_k_ho_wo_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_c_hi_wi,
const Tensor<TInWei>& wei_k_c_y_x,
Tensor<TOut>& out_n_k_ho_wo,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
DeviceMem in_n_c_hi_wi_device_buf(sizeof(TInWei) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_k_c_y_x_device_buf(sizeof(TInWei) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_n_k_ho_wo_device_buf(sizeof(TOut) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_n_c_hi_wi_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
wei_k_c_y_x_device_buf.ToDevice(wei_k_c_y_x.mData.data());
out_n_k_ho_wo_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
const auto in_n_c_hi_wi_desc = make_naive_tensor_descriptor_packed(in_n_c_hi_wi_lengths);
const auto wei_k_c_y_x_desc = make_naive_tensor_descriptor_packed(wei_k_c_y_x_lengths);
const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths);
#if 1
// cdata = 64, BlockSize = 256, 128x128x8
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlockM1 = 128;
constexpr index_t GemmNPerBlockN1 = 128;
constexpr index_t GemmKPerBlock = 8;
constexpr index_t GemmM1PerThreadM111 = 4;
constexpr index_t GemmN1PerThreadN111 = 4;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmM11N11ThreadClusterM1100 = 8;
constexpr index_t GemmM11N11ThreadClusterN1100 = 8;
constexpr index_t GemmM11N11ThreadClusterM1101 = 2;
constexpr index_t GemmM11N11ThreadClusterN1101 = 2;
using GemmABlockTransferThreadSliceLengths_K_M0_M1 = Sequence<4, 1, 1>;
using GemmABlockTransferThreadClusterLengths_K_M0_M1 = Sequence<2, 1, 128>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_K = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_M1 = 1;
using GemmBBlockTransferThreadSliceLengths_K_N0_N1 = Sequence<4, 1, 1>;
using GemmBBlockTransferThreadClusterLengths_K_N0_N1 = Sequence<2, 1, 128>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_N1 = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_N1 = 1;
constexpr index_t GemmCThreadTransferDstScalarPerVector_N11 = 1;
#endif
const auto descs =
transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw_pad(wei_k_c_y_x_desc,
in_n_c_hi_wi_desc,
out_n_k_ho_wo_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads);
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto wei_gemmk_gemmm0_gemmn1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}));
constexpr auto in_gemmk_gemmn0_gemmn1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}));
constexpr auto out_gemmm0_gemmm10_gemmm11_gemmn0_gemmn10_gemmn11_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{}));
constexpr auto wei_gemmk_gemmm0_gemmm1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0>{};
constexpr auto in_gemmk_gemmn0_gemmn1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>{};
const auto wei_gemmk_gemmm_grid_desc = descs[I0];
const auto in_gemmk_gemmn_grid_desc = descs[I1];
const auto out_gemmm_gemmn_grid_desc = descs[I2];
for(index_t i = 0; i < 5; ++i)
{
float ave_time = driver_gemm_dlops_v1r2<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum::Set,
decltype(wei_gemmk_gemmm_grid_desc),
decltype(in_gemmk_gemmn_grid_desc),
decltype(out_gemmm_gemmn_grid_desc),
GemmMPerBlockM1,
GemmNPerBlockN1,
GemmKPerBlock,
GemmM1PerThreadM111,
GemmN1PerThreadN111,
GemmKPerThread,
GemmM11N11ThreadClusterM1100,
GemmM11N11ThreadClusterN1100,
GemmM11N11ThreadClusterM1101,
GemmM11N11ThreadClusterN1101,
GemmABlockTransferThreadSliceLengths_K_M0_M1,
GemmABlockTransferThreadClusterLengths_K_M0_M1,
Sequence<2, 1, 0>, // ABlockTransferThreadClusterArrangeOrder
Sequence<2, 1, 0>, // ABlockTransferSrcAccessOrder
0, // ABlockTransferSrcVectorDim
GemmABlockTransferSrcScalarPerVector_K,
GemmABlockTransferDstScalarPerVector_M1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_K_N0_N1,
GemmBBlockTransferThreadClusterLengths_K_N0_N1,
Sequence<0, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
Sequence<0, 1, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
GemmBBlockTransferSrcScalarPerVector_N1,
GemmBBlockTransferDstScalarPerVector_N1,
false, // don't move back src coordinate after threadwise copy
Sequence<3, 4, 5, 0, 1, 2>, // CThreadTransferSrcDstAccessOrder
5, // CThreadTransferSrcDstVectorDim
GemmCThreadTransferDstScalarPerVector_N11,
decltype(wei_gemmk_gemmm0_gemmn1_grid_step_hacks),
decltype(in_gemmk_gemmn0_gemmn1_grid_step_hacks),
decltype(out_gemmm0_gemmm10_gemmm11_gemmn0_gemmn10_gemmn11_grid_step_hacks),
decltype(wei_gemmk_gemmm0_gemmm1_grid_move_slice_window_step_hacks),
decltype(in_gemmk_gemmn0_gemmn1_grid_move_slice_window_step_hacks)>(
static_cast<TInWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
wei_gemmk_gemmm_grid_desc,
in_gemmk_gemmn_grid_desc,
out_gemmm_gemmn_grid_desc,
wei_gemmk_gemmm0_gemmn1_grid_step_hacks,
in_gemmk_gemmn0_gemmn1_grid_step_hacks,
out_gemmm0_gemmm10_gemmm11_gemmn0_gemmn10_gemmn11_grid_step_hacks,
wei_gemmk_gemmm0_gemmm1_grid_move_slice_window_step_hacks,
in_gemmk_gemmn0_gemmn1_grid_move_slice_window_step_hacks,
nrepeat);
float perf = static_cast<float>(calculate_convolution_flops(
in_n_c_hi_wi_desc, wei_k_c_y_x_desc, out_n_k_ho_wo_desc)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
}
// copy result back to host
out_n_k_ho_wo_device_buf.FromDevice(out_n_k_ho_wo.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_forward_implicit_gemm_v4r4r2_dlops_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp"
#include "driver_gemm_dlops_v1r3.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_forward_implicit_gemm_v4r4r2_dlops_nhwc_kyxc_nhwk(
const InLengths& in_n_hi_wi_c_lengths,
const WeiLengths& wei_k_y_x_c_lengths,
const OutLengths& out_n_ho_wo_k_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_hi_wi_c,
const Tensor<TInWei>& wei_k_y_x_c,
Tensor<TOut>& out_n_ho_wo_k,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
DeviceMem in_n_hi_wi_c_device_buf(sizeof(TInWei) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_k_y_x_c_device_buf(sizeof(TInWei) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
#if 0
// [M, N, K0, K1] = [128, 128, 8, 1] for fp32
// cdata = 64, BlockSize = 256
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlockM1 = 128;
constexpr index_t GemmNPerBlockN1 = 128;
constexpr index_t GemmKPerBlock = 8;
constexpr index_t GemmK1 = 1;
constexpr index_t GemmM1PerThreadM111 = 4;
constexpr index_t GemmN1PerThreadN111 = 4;
constexpr index_t GemmKPerThread = 1;
using GemmM11N11ThreadClusterM110Xs = Sequence<8, 2>;
using GemmM11N11ThreadClusterN110Xs = Sequence<8, 2>;
using GemmABlockTransferThreadSliceLengths_K0_M0_M1_K1 = Sequence<4, 1, 1, 1>;
using GemmABlockTransferThreadClusterLengths_K0_M0_M1_K1 = Sequence<2, 1, 128, 1>;
using GemmABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1 = Sequence<4, 1, 1, 1>;
using GemmABlockTransferDstVectorTensorLengths_K0_M0_M1_K1 = Sequence<1, 1, 1, 1>;
using GemmBBlockTransferThreadSliceLengths_K0_N0_N1_K1 = Sequence<4, 1, 1, 1>;
using GemmBBlockTransferThreadClusterLengths_K0_N0_N1_K1 = Sequence<2, 1, 128, 1>;
using GemmBBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1 = Sequence<4, 1, 1, 1>;
using GemmBBlockTransferDstVectorTensorLengths_K0_N0_N1_K1 = Sequence<1, 1, 1, 1>;
constexpr index_t GemmCThreadTransferDstScalarPerVector_N11 = 4;
#elif 1
// [M, N, K0, K1] = [128, 128, 8, 2] for fp16
// cdata = 64, BlockSize = 256
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlockM1 = 128;
constexpr index_t GemmNPerBlockN1 = 128;
constexpr index_t GemmKPerBlock = 8;
constexpr index_t GemmK1 = 2;
constexpr index_t GemmM1PerThreadM111 = 4;
constexpr index_t GemmN1PerThreadN111 = 4;
constexpr index_t GemmKPerThread = 1;
using GemmM11N11ThreadClusterM110Xs = Sequence<8, 2>;
using GemmM11N11ThreadClusterN110Xs = Sequence<8, 2>;
using GemmABlockTransferThreadSliceLengths_K0_M0_M1_K1 = Sequence<4, 1, 1, 2>;
using GemmABlockTransferThreadClusterLengths_K0_M0_M1_K1 = Sequence<2, 1, 128, 1>;
using GemmABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1 = Sequence<4, 1, 1, 2>;
using GemmABlockTransferDstVectorTensorLengths_K0_M0_M1_K1 = Sequence<1, 1, 1, 2>;
using GemmBBlockTransferThreadSliceLengths_K0_N0_N1_K1 = Sequence<4, 1, 1, 2>;
using GemmBBlockTransferThreadClusterLengths_K0_N0_N1_K1 = Sequence<2, 1, 128, 1>;
using GemmBBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1 = Sequence<4, 1, 1, 2>;
using GemmBBlockTransferDstVectorTensorLengths_K0_N0_N1_K1 = Sequence<1, 1, 1, 2>;
constexpr index_t GemmCThreadTransferDstScalarPerVector_N11 = 4;
#elif 1
// [M, N, K0, K1] = [128, 128, 8, 4] for i8
// cdata = 64, BlockSize = 256
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlockM1 = 128;
constexpr index_t GemmNPerBlockN1 = 128;
constexpr index_t GemmKPerBlock = 8;
constexpr index_t GemmK1 = 4;
constexpr index_t GemmM1PerThreadM111 = 4;
constexpr index_t GemmN1PerThreadN111 = 4;
constexpr index_t GemmKPerThread = 1;
using GemmM11N11ThreadClusterM110Xs = Sequence<8, 2>;
using GemmM11N11ThreadClusterN110Xs = Sequence<8, 2>;
using GemmABlockTransferThreadSliceLengths_K0_M0_M1_K1 = Sequence<4, 1, 1, 4>;
using GemmABlockTransferThreadClusterLengths_K0_M0_M1_K1 = Sequence<2, 1, 128, 1>;
using GemmABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1 = Sequence<4, 1, 1, 4>;
using GemmABlockTransferDstVectorTensorLengths_K0_M0_M1_K1 = Sequence<1, 1, 1, 4>;
using GemmBBlockTransferThreadSliceLengths_K0_N0_N1_K1 = Sequence<4, 1, 1, 4>;
using GemmBBlockTransferThreadClusterLengths_K0_N0_N1_K1 = Sequence<2, 1, 128, 1>;
using GemmBBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1 = Sequence<4, 1, 1, 4>;
using GemmBBlockTransferDstVectorTensorLengths_K0_N0_N1_K1 = Sequence<1, 1, 1, 4>;
constexpr index_t GemmCThreadTransferDstScalarPerVector_N11 = 4;
#endif
const auto descs =
transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk(in_n_hi_wi_c_desc,
wei_k_y_x_c_desc,
out_n_ho_wo_k_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<GemmK1>{});
const auto in_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto wei_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto out_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto in_gemmk0_gemmm0_gemmm1_gemmk1_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 1+: GemmM0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 2+: GemmM1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}), // 3+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 1-: GemmM0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 3-: GemmM1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{})); // 3-: GemmK1
constexpr auto wei_gemmk0_gemmn0_gemmn1_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: GemmN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: GemmN1
Sequence<0, 0, 0, 0, 0, 0, 0, 0>{}), // 3+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: GemmN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: GemmN1
Sequence<0, 0, 0, 0, 0, 0, 0, 0>{})); // 3-: GemmK1
constexpr auto out_gemmm0_gemmm10_gemmm11_gemmn0_gemmn10_gemmn11_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: GemmM0
Sequence<0, 0, 0, 0, 0>{}, // 1+: GemmM10
Sequence<0, 0, 0, 0, 0>{}, // 2+: GemmM11
Sequence<0, 0, 0, 0, 0>{}, // 3+: GemmN0
Sequence<0, 0, 0, 0, 0>{}, // 4+: GemmN10
Sequence<0, 0, 0, 0, 0>{}), // 5+: GemmN11
make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0-: GemmM0
Sequence<0, 0, 0, 0, 0>{}, // 1-: GemmM10
Sequence<0, 0, 0, 0, 0>{}, // 2-: GemmM11
Sequence<0, 0, 0, 0, 0>{}, // 3-: GemmN0
Sequence<0, 0, 0, 0, 0>{}, // 4-: GemmN10
Sequence<0, 0, 0, 0, 0>{})); // 5-: GemmN11
constexpr auto in_gemmk0_gemmm0_gemmm1_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 0, 0, 0>{};
constexpr auto wei_gemmk0_gemmn0_gemmn1_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0>{};
for(index_t i = 0; i < 5; ++i)
{
float ave_time = driver_gemm_dlops_v1r3<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum::Set,
decltype(in_gemmk0_gemmm_gemmk1_grid_desc),
decltype(wei_gemmk0_gemmn_gemmk1_grid_desc),
decltype(out_gemmm_gemmn_grid_desc),
GemmMPerBlockM1,
GemmNPerBlockN1,
GemmKPerBlock,
GemmM1PerThreadM111,
GemmN1PerThreadN111,
GemmKPerThread,
GemmM11N11ThreadClusterM110Xs,
GemmM11N11ThreadClusterN110Xs,
GemmABlockTransferThreadSliceLengths_K0_M0_M1_K1,
GemmABlockTransferThreadClusterLengths_K0_M0_M1_K1,
Sequence<1, 2, 0, 3>, // ABlockTransferThreadClusterArrangeOrder
Sequence<1, 2, 0, 3>, // ABlockTransferSrcAccessOrder
GemmABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
Sequence<1, 2, 0, 3>, // ABlockTransferSrcVectorTensorContiguousDimOrder
GemmABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
GemmBBlockTransferThreadSliceLengths_K0_N0_N1_K1,
GemmBBlockTransferThreadClusterLengths_K0_N0_N1_K1,
Sequence<1, 2, 0, 3>, // BBlockTransferThreadClusterArrangeOrder
Sequence<1, 2, 0, 3>, // BBlockTransferSrcAccessOrder
GemmBBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
Sequence<1, 2, 0, 3>, // BBlockTransferSrcVectorTensorContiguousDimOrder
GemmBBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
Sequence<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
5, // CThreadTransferSrcDstVectorDim
GemmCThreadTransferDstScalarPerVector_N11,
decltype(in_gemmk0_gemmm0_gemmm1_gemmk1_grid_step_hacks),
decltype(wei_gemmk0_gemmn0_gemmn1_gemmk1_grid_step_hacks),
decltype(out_gemmm0_gemmm10_gemmm11_gemmn0_gemmn10_gemmn11_grid_step_hacks),
decltype(in_gemmk0_gemmm0_gemmm1_gemmk1_grid_move_slice_window_step_hacks),
decltype(wei_gemmk0_gemmn0_gemmn1_gemmk1_grid_move_slice_window_step_hacks)>(
static_cast<TInWei*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
in_gemmk0_gemmm_gemmk1_grid_desc,
wei_gemmk0_gemmn_gemmk1_grid_desc,
out_gemmm_gemmn_grid_desc,
in_gemmk0_gemmm0_gemmm1_gemmk1_grid_step_hacks,
wei_gemmk0_gemmn0_gemmn1_gemmk1_grid_step_hacks,
out_gemmm0_gemmm10_gemmm11_gemmn0_gemmn10_gemmn11_grid_step_hacks,
in_gemmk0_gemmm0_gemmm1_gemmk1_grid_move_slice_window_step_hacks,
wei_gemmk0_gemmn0_gemmn1_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
{
const auto N = out_n_ho_wo_k_lengths[I0];
const auto K = out_n_ho_wo_k_lengths[I3];
const auto C = wei_k_y_x_c_lengths[I3];
const auto Ho = out_n_ho_wo_k_lengths[I1];
const auto Wo = out_n_ho_wo_k_lengths[I2];
const auto Y = wei_k_y_x_c_lengths[I1];
const auto X = wei_k_y_x_c_lengths[I2];
float perf = static_cast<float>(std::size_t(2) * N * K * Ho * Wo * C * Y * X) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
// copy result back to host
out_n_ho_wo_k_device_buf.FromDevice(out_n_ho_wo_k.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp"
#include "driver_gemm_xdlops_v2r3.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw(
const InLengths& in_n_c_hi_wi_lengths,
const WeiLengths& wei_k_c_y_x_lengths,
const OutLengths& out_n_k_ho_wo_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_c_hi_wi,
const Tensor<TInWei>& wei_k_c_y_x,
Tensor<TOut>& out_n_k_ho_wo,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
DeviceMem in_n_c_hi_wi_device_buf(sizeof(TInWei) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_k_c_y_x_device_buf(sizeof(TInWei) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_n_k_ho_wo_device_buf(sizeof(TOut) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_n_c_hi_wi_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
wei_k_c_y_x_device_buf.ToDevice(wei_k_c_y_x.mData.data());
out_n_k_ho_wo_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
const auto in_n_c_hi_wi_desc = make_naive_tensor_descriptor_packed(in_n_c_hi_wi_lengths);
const auto wei_k_c_y_x_desc = make_naive_tensor_descriptor_packed(wei_k_c_y_x_lengths);
const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths);
#if 0
// [M, N, K0, K1] = [128, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [256, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif
const auto descs =
transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw_pad(wei_k_c_y_x_desc,
in_n_c_hi_wi_desc,
out_n_k_ho_wo_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<GemmK1>{});
const auto wei_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto in_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto out_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1+: GemmM
Sequence<0, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1-: GemmM
Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto in_gemmk0_gemmn_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 1+: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 1-: GemmN
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0>{};
constexpr auto in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>{};
for(index_t i = 0; i < 5; ++i)
{
float ave_time = driver_gemm_xdlops_v2r3<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum::Set,
decltype(wei_gemmk0_gemmm_gemmk1_grid_desc),
decltype(in_gemmk0_gemmn_gemmk1_grid_desc),
decltype(out_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerWave,
GemmNPerWave,
GemmK1,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmABlockTransferSrcScalarPerVector_GemmK1,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<0, 2, 1>,
Sequence<1, 0, 2>,
1,
GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
Sequence<3, 0, 1, 2, 7, 5, 4, 6>,
7,
GemmCThreadTransferDstScalarPerVector,
decltype(wei_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false>(static_cast<TInWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
wei_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmk0_gemmn_gemmk1_grid_desc,
out_gemmm_gemmn_grid_desc,
wei_gemmk0_gemmm_gemmk1_grid_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_step_hacks,
out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
float perf = static_cast<float>(calculate_convolution_flops(
in_n_c_hi_wi_desc, wei_k_c_y_x_desc, out_n_k_ho_wo_desc)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
}
// copy result back to host
out_n_k_ho_wo_device_buf.FromDevice(out_n_k_ho_wo.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp"
#include "driver_gemm_xdlops_v2r3.hpp"
#if 0
__host__ __device__ static constexpr auto
MakePaddedGridDescriptors(const AGridDesc_K0Raw_MRaw_K1& a_grid_desc_k0raw_mraw_k1,
const BGridDesc_K0Raw_NRaw_K1& b_grid_desc_k0raw_nraw_k1,
const CGridDesc_MRaw_NRaw& c_grid_desc_mraw_nraw)
{
const auto K0Raw = a_grid_desc_k0raw_mraw_k1.GetLength(I0);
const auto K1 = a_grid_desc_k0raw_mraw_k1.GetLength(I2);
const auto MRaw = c_grid_desc_mraw_nraw.GetLength(I0);
const auto NRaw = c_grid_desc_mraw_nraw.GetLength(I1);
const auto K0Pad = math::integer_least_multiple(K0Raw, K0PerBlock) - K0Raw;
const auto MPad = math::integer_least_multiple(MRaw, MPerBlock) - MRaw;
const auto NPad = math::integer_least_multiple(NRaw, NPerBlock) - NRaw;
// A
const auto a_grid_desc_k0_m_k1 = [&]() {
if constexpr(DoPad_K0 && DoPad_M)
{
return transform_tensor_descriptor(
a_grid_desc_k0_m_k1,
make_tuple(make_right_pad_transform(K0Raw, K0Pad),
make_right_pad_transform(MRaw, MPad),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
}
else if constexpr(DoPad_K0 && !DoPad_M)
{
return transform_tensor_descriptor(
a_grid_desc_k0_m_k1,
make_tuple(make_right_pad_transform(K0Raw, K0Pad),
make_pass_through_transform(MRaw),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
}
else if constexpr(!DoPad_K0 && DoPad_M)
{
return transform_tensor_descriptor(
a_grid_desc_k0_m_k1,
make_tuple(make_pass_through_transform(K0Raw),
make_right_pad_transform(MRaw, MPad),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
}
else
{
return a_grid_desc_k0raw_mraw_k1;
}
}();
// B
const auto b_grid_desc_k0_n_k1 = [&]() {
if constexpr(DoPad_K0 && DoPad_N)
{
return transform_tensor_descriptor(
b_grid_desc_k0_n_k1,
make_tuple(make_right_pad_transform(K0Raw, K0Pad),
make_right_pad_transform(NRaw, NPad),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
}
else if constexpr(DoPad_K0 && !DoPad_N)
{
return transform_tensor_descriptor(
b_grid_desc_k0_n_k1,
make_tuple(make_right_pad_transform(K0Raw, K0Pad),
make_pass_through_transform(NRaw),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
}
else if constexpr(!DoPad_K0 && DoPad_N)
{
return transform_tensor_descriptor(
b_grid_desc_k0_n_k1,
make_tuple(make_pass_through_transform(K0Raw),
make_right_pad_transform(NRaw, NPad),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
}
else
{
return b_grid_desc_k0raw_nraw_k1;
}
}();
// C
const auto c_grid_desc_m_n = [&]() {
if constexpr(DoPad_M && DoPad_N)
{
return transform_tensor_descriptor(c_grid_desc_m_n,
make_tuple(make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(DoPad_M && !DoPad_N)
{
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(!DoPad_M && DoPad_N)
{
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
reutnr c_grid_desc_m_n;
}
}();
}
#endif
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
const InLengths& in_n_hi_wi_c_lengths,
const WeiLengths& wei_k_y_x_c_lengths,
const OutLengths& out_n_ho_wo_k_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_hi_wi_c,
const Tensor<TInWei>& wei_k_y_x_c,
Tensor<TOut>& out_n_ho_wo_k,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
DeviceMem in_n_hi_wi_c_device_buf(sizeof(TInWei) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_k_y_x_c_device_buf(sizeof(TInWei) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
#if 0
// [M, N, K0, K1] = [256, 128, 4, 4], C = 128, for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 128, 4, 4], C = 128, for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [256, 256, 4, 8], C = 256, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 4;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [256, 128, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [128, 256, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 4;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 128, 4, 8], C = 64, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 64, 4, 8], C = 64, for fp16
constexpr index_t BlockSize = 128;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 64;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 32, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [128, 64, 4, 8], C = 32, for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 64;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerXDL = 32;
constexpr index_t GemmNPerXDL = 32;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 1, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif
const auto descs =
transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk(in_n_hi_wi_c_desc,
wei_k_y_x_c_desc,
out_n_ho_wo_k_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<GemmK1>{});
#if 0 // debug
const auto in_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
// HACK: hacks that control index calculation when iterating over A matrix
constexpr auto in_gemmk0_gemmm_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}, // 1+: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{}, // 1-: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>{};
#else
const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = descs[I0];
const auto GemmK0 = in_gemmk0_gemmmraw_gemmk1_grid_desc.GetLength(I0);
const auto GemmMRaw = in_gemmk0_gemmmraw_gemmk1_grid_desc.GetLength(I1);
const auto GemmMPad = math::integer_least_multiple(GemmMRaw, GemmMPerBlock) - GemmMRaw;
const auto in_gemmk0_gemmm_gemmk1_grid_desc =
transform_tensor_descriptor(in_gemmk0_gemmmraw_gemmk1_grid_desc,
make_tuple(make_pass_through_transform(GemmK0),
make_right_pad_transform(GemmMRaw, GemmMPad),
make_pass_through_transform(GemmK1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
// HACK: hacks that control index calculation when iterating over A matrix
constexpr auto in_gemmk0_gemmm_gemmk1_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 1+: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 1-: GemmM
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 0, 0, 0>{};
#endif
const auto wei_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto wei_gemmk0_gemmn_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1+: GemmN
Sequence<0, 0, 0, 0, 0>{}), // 2+: GemmK1
make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0-: GemmK0
Sequence<0, 0, 0, 0, 0>{}, // 1-: GemmN
Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0>{};
#if 0
const auto out_gemmm_gemmn_grid_desc = descs[I2];
constexpr auto out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
#else
const auto out_gemmmraw_gemmn_grid_desc = descs[I2];
const auto GemmN = out_gemmmraw_gemmn_grid_desc.GetLength(I1);
const auto out_gemmm_gemmn_grid_desc =
transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
constexpr auto out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
#endif
for(index_t i = 0; i < 5; ++i)
{
float ave_time = driver_gemm_xdlops_v2r3<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum::Set,
decltype(in_gemmk0_gemmm_gemmk1_grid_desc),
decltype(wei_gemmk0_gemmn_gemmk1_grid_desc),
decltype(out_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerXDL,
GemmNPerXDL,
GemmK1,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmABlockTransferSrcScalarPerVector_GemmK1,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmBBlockTransferSrcScalarPerVector_GemmK1,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
7,
GemmCThreadTransferDstScalarPerVector,
decltype(in_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
decltype(in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
false, // CAccessOrderMRepeatNRepeat
true, // ABlockLdsExtraM
true // BBlockLdsExtraN
>(static_cast<TInWei*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
in_gemmk0_gemmm_gemmk1_grid_desc,
wei_gemmk0_gemmn_gemmk1_grid_desc,
out_gemmm_gemmn_grid_desc,
debug::debug_driver_gemm_xdlops_v2r3::M01,
debug::debug_driver_gemm_xdlops_v2r3::N01,
in_gemmk0_gemmm_gemmk1_grid_step_hacks,
wei_gemmk0_gemmn_gemmk1_grid_step_hacks,
out_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
{
const auto N = out_n_ho_wo_k_lengths[I0];
const auto K = out_n_ho_wo_k_lengths[I3];
const auto C = wei_k_y_x_c_lengths[I3];
const auto Ho = out_n_ho_wo_k_lengths[I1];
const auto Wo = out_n_ho_wo_k_lengths[I2];
const auto Y = wei_k_y_x_c_lengths[I1];
const auto X = wei_k_y_x_c_lengths[I2];
float perf = static_cast<float>((std::size_t(2) * N * K * Ho * Wo * C * Y * X)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
// copy result back to host
out_n_ho_wo_k_device_buf.FromDevice(out_n_ho_wo_k.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "driver_convolution_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
ck::ActivTypeEnum activ_type,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1(
const InLengths& in_n_c0_hi_wi_c1_lengths,
const WeiLengths& wei_k_c0_y_x_c1_lengths,
const OutLengths& out_n_k0_ho_wo_k1_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_c0_hi_wi_c1,
const Tensor<TInWei>& wei_k_c0_y_x_c1,
const Tensor<TOut>& bias_k0_k1,
Tensor<TOut>& out_n_k0_ho_wo_k1,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
const auto N = out_n_k0_ho_wo_k1_lengths[I0];
const auto K0 = out_n_k0_ho_wo_k1_lengths[I1];
const auto Ho = out_n_k0_ho_wo_k1_lengths[I2];
const auto Wo = out_n_k0_ho_wo_k1_lengths[I3];
const auto K1 = out_n_k0_ho_wo_k1_lengths[I4];
const auto C0 = in_n_c0_hi_wi_c1_lengths[I1];
const auto Hi = in_n_c0_hi_wi_c1_lengths[I2];
const auto Wi = in_n_c0_hi_wi_c1_lengths[I3];
const auto C1 = in_n_c0_hi_wi_c1_lengths[I4];
const auto K = wei_k_c0_y_x_c1_lengths[I0];
const auto Y = wei_k_c0_y_x_c1_lengths[I2];
const auto X = wei_k_c0_y_x_c1_lengths[I3];
DeviceMem in_n_c0_hi_wi_c1_device_buf(sizeof(TInWei) *
in_n_c0_hi_wi_c1.mDesc.GetElementSpace());
DeviceMem wei_k_c0_y_x_c1_device_buf(sizeof(TInWei) * wei_k_c0_y_x_c1.mDesc.GetElementSpace());
DeviceMem bias_k0_k1_device_buf(sizeof(TOut) * bias_k0_k1.mDesc.GetElementSpace());
DeviceMem out_n_k0_ho_wo_k1_device_buf(sizeof(TOut) *
out_n_k0_ho_wo_k1.mDesc.GetElementSpace());
in_n_c0_hi_wi_c1_device_buf.ToDevice(in_n_c0_hi_wi_c1.mData.data());
wei_k_c0_y_x_c1_device_buf.ToDevice(wei_k_c0_y_x_c1.mData.data());
bias_k0_k1_device_buf.ToDevice(bias_k0_k1.mData.data());
constexpr index_t InWeiVectorSize = 8;
if(C1 % InWeiVectorSize != 0)
{
throw std::runtime_error("wrong! C1 cannot be divided by InWeiVectorSize");
}
#if 0
constexpr index_t BlockSize = 256;
constexpr index_t KPerBlock = 32;
constexpr index_t HoPerBlock = 8;
constexpr index_t WoPerBlock = 64;
constexpr index_t E1 = C0 * 9;
constexpr index_t E2 = 1;
constexpr index_t E1PerBlock = C0;
constexpr index_t KPerThread = 16;
constexpr index_t HoPerThread = 2;
constexpr index_t WoPerThread = 2;
constexpr index_t EPerThread = 1;
using ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2 = Sequence<1, 9, 1, E2>;
using ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2 = Sequence<1, E1PerBlock, KPerBlock, 1>;
constexpr index_t ABlockTransferSrcScalarPerVector_E2 = E2;
constexpr index_t ABlockTransferDstScalarPerVector_E2 = E2;
constexpr index_t BThreadTransferSrcScalarPerVector_E2 = E2;
constexpr index_t CThreadTransferDstScalarPerVector_K = K1;
#elif 1
constexpr index_t BlockSize = 64;
constexpr index_t KPerBlock = 8;
constexpr index_t HoPerBlock = 8;
constexpr index_t WoPerBlock = 32;
constexpr index_t E1 = 2 * 9;
constexpr index_t E2 = 1;
constexpr index_t K2 = 2;
constexpr index_t E1PerBlock = 2;
constexpr index_t KPerThread = KPerBlock;
constexpr index_t HoPerThread = 2;
constexpr index_t WoPerThread = 2;
constexpr index_t EPerThread = 1;
using ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2 = Sequence<1, 9, 1, 1, E2>;
using ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2 =
Sequence<1, E1PerBlock, 1, KPerBlock, 1>;
constexpr index_t ABlockTransferSrcScalarPerVector_E2 = E2;
constexpr index_t ABlockTransferDstScalarPerVector_E2 = E2;
constexpr index_t BThreadTransferSrcScalarPerVector_E2 = E2;
constexpr index_t CThreadTransferDstScalarPerVector_K = InWeiVectorSize;
#endif
if(KPerThread % InWeiVectorSize != 0)
{
throw std::runtime_error("wrong! C1 cannot be divided by InWeiVectorSize");
}
const auto in_n_c0_hi_wi_c1_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, C0, Hi, Wi, E2));
const auto wei_k_c0_y_x_c1_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, C0, Y, X, E2));
const auto out_n_k0_ho_wo_k1_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, K0, Ho, Wo, K1));
constexpr auto conv_driver =
DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nc0hwc1_kc0yxc1_nk0hwk1_outpad<
BlockSize,
typename vector_type<TInWei, InWeiVectorSize>::type,
TAcc,
TOut,
E1,
E2,
K2,
KPerBlock,
HoPerBlock,
WoPerBlock,
E1PerBlock,
KPerThread,
HoPerThread,
WoPerThread,
EPerThread,
ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2,
ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2,
ABlockTransferSrcScalarPerVector_E2,
ABlockTransferDstScalarPerVector_E2,
BThreadTransferSrcScalarPerVector_E2,
CThreadTransferDstScalarPerVector_K,
activ_type>{};
std::cerr << "conv_bias_activ_input_"
<< "n" << N << "c" << C0 << "h" << Hi << "w" << Wi << "c" << C1 << "_filter_k" << K
<< "c" << C0 << "y" << Y << "x" << X << "c" << C1 << "_convout_n" << N << "k" << K0
<< "h" << Ho << "w" << Wo << "k" << K1 << std::endl;
for(int i = 0; i < 5; i++)
{
const auto ave_time =
conv_driver.Run(wei_k_c0_y_x_c1_desc,
in_n_c0_hi_wi_c1_desc,
out_n_k0_ho_wo_k1_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
wei_k_c0_y_x_c1_device_buf.GetDeviceBuffer()),
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
in_n_c0_hi_wi_c1_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(bias_k0_k1_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_k0_ho_wo_k1_device_buf.GetDeviceBuffer()),
nrepeat);
{
float perf = static_cast<float>(std::size_t(2) * N * K * Ho * Wo * C0 * C1 * Y * X) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
out_n_k0_ho_wo_k1_device_buf.FromDevice(out_n_k0_ho_wo_k1.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_forward_implicit_gemm_v6r1_dlops_nchw_kcyx_nkhw.hpp deleted 100644 → 0
View file @ cc50b687
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v6r1_nchw_kcyx_nkhw.hpp"
#include "driver_contraction_dlops_v1r2.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_forward_implicit_gemm_v6r1_dlops_nchw_kcyx_nkhw(
const InLengths& in_n_c_hi_wi_lengths,
const WeiLengths& wei_k_c_y_x_lengths,
const OutLengths& out_n_k_ho_wo_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_c_hi_wi,
const Tensor<TInWei>& wei_k_c_y_x,
Tensor<TOut>& out_n_k_ho_wo,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
DeviceMem in_n_c_hi_wi_device_buf(sizeof(TInWei) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_k_c_y_x_device_buf(sizeof(TInWei) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_n_k_ho_wo_device_buf(sizeof(TOut) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_n_c_hi_wi_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
wei_k_c_y_x_device_buf.ToDevice(wei_k_c_y_x.mData.data());
out_n_k_ho_wo_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
const auto in_desc_n_c_hi_wi = make_naive_tensor_descriptor_packed(in_n_c_hi_wi_lengths);
const auto wei_desc_k_c_y_x = make_naive_tensor_descriptor_packed(wei_k_c_y_x_lengths);
const auto out_desc_n_k_ho_wo = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths);
#if 1
// [8, 1, 128, 1] * [8, 4, 32, 1] = [1, 128, 4, 32] for fp32
// cdata = 64, BlockSize = 256
constexpr index_t BlockSize = 256;
constexpr index_t GN0 = 4;
constexpr index_t GK1 = 1;
constexpr index_t GM1PerBlockGM11 = 128;
constexpr index_t GN1PerBlockGN11 = 32;
constexpr index_t GK0PerBlock = 8;
constexpr index_t BM1PerThreadBM11 = 4;
constexpr index_t BN1PerThreadBN11 = 4;
constexpr index_t BK0PerThread = 1;
using BM10BN10ThreadClusterBM10Xs = Sequence<8, 2>;
using BM10BN10ThreadClusterBN10Xs = Sequence<8, 2>;
using ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1 = Sequence<4, 1, 1, 1, 1>;
using ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1 = Sequence<2, 1, 1, 128, 1>;
using ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1 = Sequence<4, 1, 1, 1, 1>;
using ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1 = Sequence<1, 1, 1, 1, 1>;
using BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1 = Sequence<1, 4, 1, 1, 1>;
using BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1 = Sequence<8, 1, 1, 32, 1>;
using BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1 = Sequence<1, 1, 1, 1, 1>;
using BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1 = Sequence<1, 1, 1, 1, 1>;
constexpr index_t CThreadTransferDstScalarPerVector_BN1 = 1;
#elif 1
// [8, 1, 128, 2] * [8, 4, 32, 2] = [1, 128, 4, 32] for fp16
// cdata = 64, BlockSize = 256
constexpr index_t BlockSize = 256;
constexpr index_t GN0 = 4;
constexpr index_t GK1 = 2;
constexpr index_t GM1PerBlockGM11 = 128;
constexpr index_t GN1PerBlockGN11 = 32;
constexpr index_t GK0PerBlock = 8;
constexpr index_t BM1PerThreadBM11 = 4;
constexpr index_t BN1PerThreadBN11 = 4;
constexpr index_t BK0PerThread = 1;
using BM10BN10ThreadClusterBM10Xs = Sequence<8, 2>;
using BM10BN10ThreadClusterBN10Xs = Sequence<8, 2>;
using ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1 = Sequence<4, 1, 1, 1, 2>;
using ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1 = Sequence<2, 1, 1, 128, 1>;
using ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1 = Sequence<4, 1, 1, 1, 1>;
using ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1 = Sequence<1, 1, 1, 1, 2>;
using BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1 = Sequence<1, 4, 1, 1, 2>;
using BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1 = Sequence<8, 1, 1, 32, 1>;
using BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1 = Sequence<1, 1, 1, 1, 1>;
using BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1 = Sequence<1, 1, 1, 1, 2>;
constexpr index_t CThreadTransferDstScalarPerVector_BN1 = 1;
#endif
const auto descs =
transform_forward_convolution_into_contraction_v6r1_nchw_kcyx_nkhw_pad(wei_desc_k_c_y_x,
in_desc_n_c_hi_wi,
out_desc_n_k_ho_wo,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<GN0>{},
Number<GK1>{});
const auto wei_grid_desc_gk0_gm0_gm1_gk1 = descs[I0];
const auto in_grid_desc_gk0_gn0_gn1_gk1 = descs[I1];
const auto out_grid_desc_gm0_gm1_gn0_gn1 = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto wei_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 0+: GK0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 1+: GM0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 2+: GM10
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 3+: GM11
Sequence<0, 0, 0, 0, 0, 0, 0>{}), // 4+: GK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 0-: GK0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 1-: GM0
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 2-: GM10
Sequence<0, 0, 0, 0, 0, 0, 0>{}, // 3-: GM11
Sequence<0, 0, 0, 0, 0, 0, 0>{})); // 4-: GK1
constexpr auto in_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: GK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 1+: GN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 2+: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 3+: GN11
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 4+: GK1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: GK0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 1-: GN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 2-: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 3-: GN11
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 4-: GK1
constexpr auto out_grid_step_hacks = make_tuple(
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: GM10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}, // 1+: BM0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}, // 2+: BM1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}, // 4+: BN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}), // 5+: GN1
make_tuple(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: GM10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{}, // 1-: BM0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{}, // 2-: BM1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: GN10
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{}, // 4-: BN0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{})); // 5-: GN1
constexpr auto wei_grid_move_slice_window_step_hacks = Sequence<0, 0, 0, 0, 0, 0, 0>{};
constexpr auto in_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 2, 0, 0, 0, 0, 0>{};
for(index_t i = 0; i < 5; ++i)
{
float ave_time = driver_contraction_dlops_v1r2<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum::Set,
decltype(wei_grid_desc_gk0_gm0_gm1_gk1),
decltype(in_grid_desc_gk0_gn0_gn1_gk1),
decltype(out_grid_desc_gm0_gm1_gn0_gn1),
GM1PerBlockGM11,
GN1PerBlockGN11,
GK0PerBlock,
BM1PerThreadBM11,
BN1PerThreadBN11,
BK0PerThread,
BM10BN10ThreadClusterBM10Xs,
BM10BN10ThreadClusterBN10Xs,
ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1,
ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1,
Sequence<1, 2, 3, 0, 4>, // ABlockTransferThreadClusterArrangeOrder
Sequence<3, 2, 1, 0, 4>, // ABlockTransferSrcAccessOrder
ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1,
ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1,
Sequence<0, 1, 2, 3, 4>, // ABlockTransferSrcVectorTensorContiguousDimOrder
BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1,
BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1,
Sequence<0, 4, 1, 2, 3>, // BBlockTransferThreadClusterArrangeOrder
Sequence<4, 3, 2, 0, 1>, // BBlockTransferSrcAccessOrder
BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1,
BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1,
Sequence<0, 1, 2, 3, 4>, // BBlockTransferSrcVectorTensorContiguousDimOrder
Sequence<3, 4, 5, 0, 1, 2>, // CThreadTransferSrcDstAccessOrder
5, // CThreadTransferSrcDstVectorDim
CThreadTransferDstScalarPerVector_BN1,
decltype(wei_grid_step_hacks),
decltype(in_grid_step_hacks),
decltype(out_grid_step_hacks),
decltype(wei_grid_move_slice_window_step_hacks),
decltype(in_grid_move_slice_window_step_hacks)>(
static_cast<TInWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
wei_grid_desc_gk0_gm0_gm1_gk1,
in_grid_desc_gk0_gn0_gn1_gk1,
out_grid_desc_gm0_gm1_gn0_gn1,
wei_grid_step_hacks,
in_grid_step_hacks,
out_grid_step_hacks,
wei_grid_move_slice_window_step_hacks,
in_grid_move_slice_window_step_hacks,
nrepeat);
float perf = static_cast<float>(calculate_convolution_flops(
in_desc_n_c_hi_wi, wei_desc_k_c_y_x, out_desc_n_k_ho_wo)) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
}
// copy result back to host
out_n_k_ho_wo_device_buf.FromDevice(out_n_k_ho_wo.mData.data());
}
library/include/ck/library/obselete_driver_offline/device_convolution_maxpool_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp deleted 100644 → 0
View file @ cc50b687
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "driver_convolution_maxpool_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
ck::ActivTypeEnum activ_type,
typename InLengths,
typename WeiLengths,
typename MaxLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_maxpool_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1(
const InLengths& in_n_c0_hi_wi_c1_lengths,
const WeiLengths& wei_k_c0_y_x_c1_lengths,
const MaxLengths& max_n_k0_hx_wx_k1_lengths,
const OutLengths& out_n_k0_ho_wo_k1_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_c0_hi_wi_c1,
const Tensor<TInWei>& wei_k_c0_y_x_c1,
const Tensor<TOut>& bias_k0_k1,
Tensor<TOut>& out_n_k0_ho_wo_k1,
Tensor<TOut>& max_n_k0_hx_wx_k1,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
const auto N = out_n_k0_ho_wo_k1_lengths[I0];
const auto K0 = out_n_k0_ho_wo_k1_lengths[I1];
const auto Ho = out_n_k0_ho_wo_k1_lengths[I2];
const auto Wo = out_n_k0_ho_wo_k1_lengths[I3];
const auto K1 = out_n_k0_ho_wo_k1_lengths[I4];
const auto C0 = in_n_c0_hi_wi_c1_lengths[I1];
const auto Hi = in_n_c0_hi_wi_c1_lengths[I2];
const auto Wi = in_n_c0_hi_wi_c1_lengths[I3];
const auto C1 = in_n_c0_hi_wi_c1_lengths[I4];
const auto K = wei_k_c0_y_x_c1_lengths[I0];
const auto Y = wei_k_c0_y_x_c1_lengths[I2];
const auto X = wei_k_c0_y_x_c1_lengths[I3];
const auto Hx = max_n_k0_hx_wx_k1_lengths[I2];
const auto Wx = max_n_k0_hx_wx_k1_lengths[I3];
DeviceMem in_n_c0_hi_wi_c1_device_buf(sizeof(TInWei) *
in_n_c0_hi_wi_c1.mDesc.GetElementSpace());
DeviceMem wei_k_c0_y_x_c1_device_buf(sizeof(TInWei) * wei_k_c0_y_x_c1.mDesc.GetElementSpace());
DeviceMem bias_k0_k1_device_buf(sizeof(TOut) * bias_k0_k1.mDesc.GetElementSpace());
DeviceMem out_n_k0_ho_wo_k1_device_buf(sizeof(TOut) *
out_n_k0_ho_wo_k1.mDesc.GetElementSpace());
DeviceMem max_n_k0_hx_wx_k1_device_buf(sizeof(TOut) *
max_n_k0_hx_wx_k1.mDesc.GetElementSpace());
in_n_c0_hi_wi_c1_device_buf.ToDevice(in_n_c0_hi_wi_c1.mData.data());
wei_k_c0_y_x_c1_device_buf.ToDevice(wei_k_c0_y_x_c1.mData.data());
bias_k0_k1_device_buf.ToDevice(bias_k0_k1.mData.data());
max_n_k0_hx_wx_k1_device_buf.ToDevice(max_n_k0_hx_wx_k1.mData.data());
constexpr index_t InWeiVectorSize = 8;
if(C1 % InWeiVectorSize != 0)
{
throw std::runtime_error("wrong! C1 cannot be divided by InWeiVectorSize");
}
#if 0
constexpr index_t BlockSize = 256;
constexpr index_t KPerBlock = 32;
constexpr index_t HoPerBlock = 8;
constexpr index_t WoPerBlock = 64;
constexpr index_t E1 = C0 * 9;
constexpr index_t E2 = 1;
constexpr index_t E1PerBlock = C0;
constexpr index_t KPerThread = 16;
constexpr index_t HoPerThread = 2;
constexpr index_t WoPerThread = 2;
constexpr index_t EPerThread = 1;
using ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2 = Sequence<1, 9, 1, E2>;
using ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2 = Sequence<1, E1PerBlock, KPerBlock, 1>;
constexpr index_t ABlockTransferSrcScalarPerVector_E2 = E2;
constexpr index_t ABlockTransferDstScalarPerVector_E2 = E2;
constexpr index_t BThreadTransferSrcScalarPerVector_E2 = E2;
constexpr index_t CThreadTransferDstScalarPerVector_K = K1;
#elif 1
constexpr index_t BlockSize = 64;
constexpr index_t KPerBlock = 8;
constexpr index_t HoPerBlock = 8;
constexpr index_t WoPerBlock = 32;
constexpr index_t E1 = 2 * 9;
constexpr index_t E2 = 1;
constexpr index_t K2 = 2;
constexpr index_t E1PerBlock = 2;
constexpr index_t KPerThread = KPerBlock;
constexpr index_t HoPerThread = 2;
constexpr index_t WoPerThread = 2;
constexpr index_t EPerThread = 1;
using ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2 = Sequence<1, 9, 1, 1, E2>;
using ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2 =
Sequence<1, E1PerBlock, 1, KPerBlock, 1>;
constexpr index_t ABlockTransferSrcScalarPerVector_E2 = E2;
constexpr index_t ABlockTransferDstScalarPerVector_E2 = E2;
constexpr index_t BThreadTransferSrcScalarPerVector_E2 = E2;
constexpr index_t CThreadTransferDstScalarPerVector_K = InWeiVectorSize;
#endif
if(KPerThread % InWeiVectorSize != 0)
{
throw std::runtime_error("wrong! C1 cannot be divided by InWeiVectorSize");
}
const auto in_n_c0_hi_wi_c1_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, C0, Hi, Wi, E2));
const auto wei_k_c0_y_x_c1_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, C0, Y, X, E2));
const auto max_n_k0_hx_wx_k1_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, K0, Hx, Wx, K1));
const auto out_n_k0_ho_wo_k1_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, K0, Ho, Wo, K1));
constexpr auto conv_driver =
DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nc0hwc1_kc0yxc1_nk0hwk1_maxpool<
BlockSize,
typename vector_type<TInWei, InWeiVectorSize>::type,
TAcc,
TOut,
E1,
E2,
K2,
KPerBlock,
HoPerBlock,
WoPerBlock,
E1PerBlock,
KPerThread,
HoPerThread,
WoPerThread,
EPerThread,
ABlockTransferThreadSliceLengths_E0_E1_K0_K1_E2,
ABlockTransferThreadClusterLengths_E0_E1_K0_K1_E2,
ABlockTransferSrcScalarPerVector_E2,
ABlockTransferDstScalarPerVector_E2,
BThreadTransferSrcScalarPerVector_E2,
CThreadTransferDstScalarPerVector_K,
activ_type>{};
std::cerr << "conv_bias_activ_maxpool_input_"
<< "n" << N << "c" << C0 << "h" << Hi << "w" << Wi << "c" << C1 << "_filter_k" << K
<< "c" << C0 << "y" << Y << "x" << X << "c" << C1 << "_convout_n" << N << "k" << K0
<< "h" << Ho << "w" << Wo << "k" << K1 << "_maxpoolout_n" << N << "k" << K0 << "h"
<< Ho / 2 << "w" << Wo / 2 << "k" << K1 << std::endl;
for(int i = 0; i < 5; i++)
{
const auto ave_time =
conv_driver.Run(wei_k_c0_y_x_c1_desc,
in_n_c0_hi_wi_c1_desc,
out_n_k0_ho_wo_k1_desc,
max_n_k0_hx_wx_k1_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
wei_k_c0_y_x_c1_device_buf.GetDeviceBuffer()),
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
in_n_c0_hi_wi_c1_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(bias_k0_k1_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_k0_ho_wo_k1_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(max_n_k0_hx_wx_k1_device_buf.GetDeviceBuffer()),
nrepeat);
{
float perf = static_cast<float>(std::size_t(2) * N * K * Ho * Wo * C0 * C1 * Y * X) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
<< std::endl;
}
}
out_n_k0_ho_wo_k1_device_buf.FromDevice(out_n_k0_ho_wo_k1.mData.data());
max_n_k0_hx_wx_k1_device_buf.FromDevice(max_n_k0_hx_wx_k1.mData.data());
}
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