Skip to content

GitLab

Unverified Commit 3835318c authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

xdlops_v4r4_fwd fp32/fp16 (#34) 



* create files for xdlops

* working on blockwise_gemm_xdlops

* add KReduction

* add m/n repeats

* add 2x2 pipeline

* added 128x128 wavegemm

* use StaticBuffer of vector_type

* break vector type to blk_size

* add kpack into xldops_gemm and blockwise_gemm

* abroadcast only

* add fp32 mfma instructions

* adding fp16 mfma

* pack half4_t

* rename kperwave to kpack

* add 32x32x8fp16

* add fp16 mfma

* clean code

* clean code

* V4r4 xdlops kpack (#35)

* add kpack with incorrect results

* bug fix for make_dynamic_naive_tensor_descriptor_aligned_v2

* add 1x1 kernel

* add gridwise_gemm_v2 - single_buffer

* enabled dwordx4 for fp16
Co-authored-by: default avatarChao Liu <chao.liu2@amd.com>

* refactor fwd-v4r4-xdlops

* add v4r4-nhwc-xdlop

* improve some perf of nhwc and nchw by tuning parameters, and change scheuduling in gridwise-gemm loop

* tweak scheduling in gridwise gemm

* add v4r3 with a single output copy

* init commit: output with slice win

* adding sliceWin

* add multiple repeats pattern

* starting adding bwd-v4r1-xdlops

* use tuple as SrcBuffer

* adding bwd-data v4r1 nhwc xdlops

* fix bug in make_dynamic_naive_tensor_descriptor_aligned_v2()

* fix bug in host bwd-data conv

* initial implementation of bwd-data v4r1 nhwc xdlops

* add launch bound flags

* enable launch bound

* add m/nrepeat=4

* tweak bwd-data v4r1 nhwc xdlops

* added bwd-data v4r1 nhwc xlops with output A and weight B

* add fwd-v4r4 nhwc xdlops, A input, B weight, C output
Co-authored-by: default avatarChao Liu <chao.liu2@amd.com>
parent 1685048a
Hide whitespace changes
Inline Side-by-side
Showing with 1648 additions and 111 deletions
driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4_xdlops_nchw_kcyx_nkhw.hpp 0 → 100644
View file @ 3835318c
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "driver_dynamic_convolution_forward_implicit_gemm_v4r4_xdlops_nchw_kcyx_nkhw.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_dynamic_convolution_forward_implicit_gemm_v4r4_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>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
constexpr auto I6 = Number<6>{};
constexpr auto I7 = Number<7>{};
constexpr auto I8 = Number<8>{};
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_dynamic_naive_tensor_descriptor_packed_v2(in_n_c_hi_wi_lengths);
const auto wei_k_c_y_x_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(wei_k_c_y_x_lengths);
const auto out_n_k_ho_wo_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(out_n_k_ho_wo_lengths);
#if 0
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmKPack = 8;
constexpr index_t MRepeat = 1;
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_GemmK = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_KPack = 8;
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_KPack = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 1;
#elif 0
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmKPack = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_KPack = 8;
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_KPack = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 1;
#elif 0
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmKPack = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_KPack = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 2>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_KPack = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 1;
#elif 1
// [M, N, K0, K1] = [256, 128, 4, 4]
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmKPack = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_KPack = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_KPack = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 1;
#elif 1
// [M, N, K0, K1] = [128, 128, 4, 4]
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmKPack = 4;
constexpr index_t MRepeat = 1;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_KPack = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_KPack = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 1;
#endif
const auto descs =
#if 1
transform_forward_convolution_into_gemm_v4r4_xdlops_nchw_kcyx_nkhw_pad
#else
transform_forward_convolution_into_gemm_v4r4_xdlops_nchw_kcyx_nkhw_1x1
#endif
<TInWei, GemmMPerBlock, GemmNPerBlock, GemmMPerWave, GemmNPerWave, GemmKPack>(
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);
for(index_t i = 0; i < 5; ++i)
{
#if 0
float ave_time = launch_kernel_dynamic_gemm_xdlops_v1
#else
float ave_time = launch_kernel_dynamic_gemm_xdlops_v2
#endif
<BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperation::Set,
decltype(descs[I0]),
decltype(descs[I1]),
decltype(descs[I2]),
decltype(descs[I3]),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerWave,
GemmNPerWave,
GemmKPack,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmABlockTransferSrcScalarPerVector_GemmK,
GemmABlockTransferDstScalarPerVector_KPack,
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_KPack,
false, // don't move back src coordinate after threadwise copy, which will be fused
// with MoveSrcSliceWindow() to save addr computation
Sequence<2, 3, 0, 1>,
3,
GemmCThreadTransferDstScalarPerVector_GemmN1,
decltype(descs[I4]),
decltype(descs[I5]),
decltype(descs[I6]),
decltype(descs[I7]),
decltype(descs[I8])>(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()),
descs[I0],
descs[I1],
descs[I2],
descs[I3],
descs[I4],
descs[I5],
descs[I6],
descs[I7],
descs[I8],
nrepeat);
float perf = (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());
}
driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp 0 → 100644
View file @ 3835318c
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp"
#include "driver_dynamic_gemm_xdlops_v2r2.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_dynamic_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>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
constexpr auto I6 = Number<6>{};
constexpr auto I7 = Number<7>{};
constexpr auto I8 = Number<8>{};
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_dynamic_naive_tensor_descriptor_packed_v2(in_n_c_hi_wi_lengths);
const auto wei_k_c_y_x_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(wei_k_c_y_x_lengths);
const auto out_n_k_ho_wo_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(out_n_k_ho_wo_lengths);
#if 0
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 1;
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, 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;
#elif 0
// [M, N, K0, K1] = [256, 128, 4, 8]
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
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, 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;
#elif 0
// [M, N, K0, K1] = [256, 128, 4, 8]
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
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;
#elif 1
// [M, N, K0, K1] = [256, 128, 4, 4]
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
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 = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
// [M, N, K0, K1] = [128, 128, 4, 4]
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 1;
constexpr index_t NRepeat = 1;
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 = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
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_iterator_hacks = make_tuple(
make_tuple(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>{}));
constexpr auto in_gemmk0_gemmn_gemmk1_grid_iterator_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, 1, 0, 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, 2, 0, 0, 0>{}));
constexpr auto out_m0_m1_m2_n_grid_iterator_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>{}),
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>{}));
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks =
Sequence<0, 0, 0, 0, 0>{};
constexpr auto in_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_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_dynamic_gemm_xdlops_v2r2<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperation::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>,
3,
GemmCThreadTransferDstScalarPerVector,
decltype(wei_gemmk0_gemmm_gemmk1_grid_iterator_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_iterator_hacks),
decltype(out_m0_m1_m2_n_grid_iterator_hacks),
decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_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_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmk0_gemmn_gemmk1_grid_desc,
out_gemmm_gemmn_grid_desc,
wei_gemmk0_gemmm_gemmk1_grid_iterator_hacks,
in_gemmk0_gemmn_gemmk1_grid_iterator_hacks,
out_m0_m1_m2_n_grid_iterator_hacks,
wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks,
in_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_hacks,
nrepeat);
float perf = (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());
}
driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk.hpp 0 → 100644
View file @ 3835318c
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp"
#include "driver_dynamic_gemm_xdlops_v2r2.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_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>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
constexpr auto I6 = Number<6>{};
constexpr auto I7 = Number<7>{};
constexpr auto I8 = Number<8>{};
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_dynamic_naive_tensor_descriptor_packed_v2(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(out_n_ho_wo_k_lengths);
#if 1
// [M, N, K0, K1] = [256, 128, 4, 8]
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
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 = 4;
#endif
const auto descs =
transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk_pad(wei_k_y_x_c_desc,
in_n_hi_wi_c_desc,
out_n_ho_wo_k_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_iterator_hacks = make_tuple(
make_tuple(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>{}));
constexpr auto in_gemmk0_gemmn_gemmk1_grid_iterator_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, 1, 0, 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, 2, 0, 0, 0>{}));
constexpr auto out_m0_m1_m2_n_grid_iterator_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>{}),
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>{}));
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks =
Sequence<0, 0, 0, 0, 0>{};
constexpr auto in_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_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_dynamic_gemm_xdlops_v2r2<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperation::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,
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>,
2,
GemmCThreadTransferDstScalarPerVector,
decltype(wei_gemmk0_gemmm_gemmk1_grid_iterator_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_iterator_hacks),
decltype(out_m0_m1_m2_n_grid_iterator_hacks),
decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_hacks)>(
static_cast<TInWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_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_iterator_hacks,
in_gemmk0_gemmn_gemmk1_grid_iterator_hacks,
out_m0_m1_m2_n_grid_iterator_hacks,
wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks,
in_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_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 Hi = in_n_hi_wi_c_lengths[I1];
const auto Wi = in_n_hi_wi_c_lengths[I2];
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 = (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());
}
driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r3_xdlops_nhwc_kyxc_nhwk.hpp 0 → 100644
View file @ 3835318c
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp"
#include "driver_dynamic_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_dynamic_convolution_forward_implicit_gemm_v4r4r3_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>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
constexpr auto I6 = Number<6>{};
constexpr auto I7 = Number<7>{};
constexpr auto I8 = Number<8>{};
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_dynamic_naive_tensor_descriptor_packed_v2(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(out_n_ho_wo_k_lengths);
#if 0
// [M, N, K0, K1] = [256, 256, 4, 8] 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 GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 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 = 4;
#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_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#elif 0
// [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 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_GemmK1 = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#endif
const auto descs =
transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk_pad(wei_k_y_x_c_desc,
in_n_hi_wi_c_desc,
out_n_ho_wo_k_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_iterator_hacks = make_tuple(
make_tuple(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>{}));
constexpr auto in_gemmk0_gemmn_gemmk1_grid_iterator_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, 1, 0, 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, 2, 0, 0, 0>{}));
constexpr auto out_m0_m1_m2_n_grid_iterator_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{}));
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks =
Sequence<0, 0, 0, 0, 0>{};
constexpr auto in_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_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_dynamic_gemm_xdlops_v2r3<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperation::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,
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>,
6,
GemmCThreadTransferDstScalarPerVector,
decltype(wei_gemmk0_gemmm_gemmk1_grid_iterator_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_iterator_hacks),
decltype(out_m0_m1_m2_n_grid_iterator_hacks),
decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_hacks),
false // CAccessOrderMRepeatNRepeat
>(static_cast<TInWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_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_iterator_hacks,
in_gemmk0_gemmn_gemmk1_grid_iterator_hacks,
out_m0_m1_m2_n_grid_iterator_hacks,
wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks,
in_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_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 Hi = in_n_hi_wi_c_lengths[I1];
const auto Wi = in_n_hi_wi_c_lengths[I2];
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 = (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());
}
driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp 0 → 100644
View file @ 3835318c
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp"
#include "driver_dynamic_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_dynamic_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>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
constexpr auto I6 = Number<6>{};
constexpr auto I7 = Number<7>{};
constexpr auto I8 = Number<8>{};
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_dynamic_naive_tensor_descriptor_packed_v2(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(out_n_ho_wo_k_lengths);
#if 0
// [M, N, K0, K1] = [256, 256, 4, 8] 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 GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 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] 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_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
// [M, N, K0, K1] = [128, 256, 4, 8] 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_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
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 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_GemmK1 = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
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 = 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_GemmK1 = 8;
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 = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
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;
#endif
const auto descs =
transform_forward_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 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_gemmm_gemmk1_grid_iterator_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 wei_gemmk0_gemmn_gemmk1_grid_iterator_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 out_m0_m1_m2_n_grid_iterator_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: MRepeat
Sequence<0, 0, 0, 0, 0>{}, // 1+: NRepeat
Sequence<0, 0, 0, 0, 0>{}, // 2+: MWaves
Sequence<0, 0, 0, 0, 0>{}, // 3+: NWaves
Sequence<0, 0, 0, 0, 0>{}, // 4+: M0
Sequence<0, 0, 0, 0, 0>{}, // 5+: M1
Sequence<0, 0, 0, 0, 0>{}, // 6+: M2
Sequence<0, 0, 0, 0, 0>{}), // 7+: N1
make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0-: MRepeat
Sequence<0, 0, 0, 0, 0>{}, // 1-: NRepeat
Sequence<0, 0, 0, 0, 0>{}, // 2-: MWaves
Sequence<0, 0, 0, 0, 0>{}, // 3-: NWaves
Sequence<0, 0, 0, 0, 0>{}, // 4-: M0
Sequence<0, 0, 0, 0, 0>{}, // 5-: M1
Sequence<0, 0, 0, 0, 0>{}, // 6-: M2
Sequence<0, 0, 0, 0, 0>{})); // 7-: N1
constexpr auto wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_hacks =
Sequence<0, 0, 0, 0, 0>{};
constexpr auto in_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_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_dynamic_gemm_xdlops_v2r3<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperation::Set,
decltype(in_gemmk0_gemmm_gemmk1_grid_desc),
decltype(wei_gemmk0_gemmn_gemmk1_grid_desc),
decltype(out_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerWave,
GemmNPerWave,
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_iterator_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_iterator_hacks),
decltype(out_m0_m1_m2_n_grid_iterator_hacks),
decltype(in_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks),
decltype(wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_hacks),
false // CAccessOrderMRepeatNRepeat
>(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_gemmm_gemmk1_grid_iterator_hacks,
wei_gemmk0_gemmn_gemmk1_grid_iterator_hacks,
out_m0_m1_m2_n_grid_iterator_hacks,
in_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks,
wei_gemmk0_gemmn_gemmk1_grid_move_slice_window_iterator_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 Hi = in_n_hi_wi_c_lengths[I1];
const auto Wi = in_n_hi_wi_c_lengths[I2];
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 = (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());
}
driver/include/host_conv.hpp
View file @ 3835318c
#pragma once #pragma once
#include "host_tensor.hpp" #include "host_tensor.hpp"
template <class TIn, template <typename TIn,
class TWei, typename TWei,
class TOut, typename TOut,
class ConvStrides, typename ConvStrides,
class ConvDilations, typename ConvDilations,
class InLeftPads, typename InLeftPads,
class InRightPads> typename InRightPads>
void host_direct_convolution(const Tensor<TIn>& in, void host_direct_convolution(const Tensor<TIn>& in,
const Tensor<TWei>& wei, const Tensor<TWei>& wei,
Tensor<TOut>& out, Tensor<TOut>& out,
...@@ -88,7 +88,7 @@ void host_direct_convolution(const Tensor<TIn>& in, ...@@ -88,7 +88,7 @@ void host_direct_convolution(const Tensor<TIn>& in,
} }
} }
template <class TIn, class TWei, class TOut, class InLeftPads, class InRightPads> template <typename TIn, typename TWei, typename TOut, typename InLeftPads, typename InRightPads>
void host_winograd_3x3_convolution(const Tensor<TIn>& in_nchw, void host_winograd_3x3_convolution(const Tensor<TIn>& in_nchw,
const Tensor<TWei>& wei_kcyx, const Tensor<TWei>& wei_kcyx,
Tensor<TOut>& out_nkhw, Tensor<TOut>& out_nkhw,
......
driver/include/host_conv_bwd_data.hpp
View file @ 3835318c
...@@ -6,56 +6,62 @@ template <typename TIn, ...@@ -6,56 +6,62 @@ template <typename TIn,
typename TOut, typename TOut,
typename ConvStrides, typename ConvStrides,
typename ConvDilations, typename ConvDilations,
typename LeftPads, typename InLeftPads,
typename RightPads> typename InRightPads>
void host_direct_convolution_backward_data(Tensor<TIn>& in_nchw, void host_direct_convolution_backward_data(Tensor<TIn>& in,
const Tensor<TWei>& wei_kcyx, const Tensor<TWei>& wei,
const Tensor<TOut>& out_nkhw, const Tensor<TOut>& out,
ConvStrides, const ConvStrides& conv_strides,
ConvDilations, const ConvDilations& conv_dilations,
LeftPads, const InLeftPads& in_left_pads,
RightPads) const InRightPads& in_right_pads,
const ConvTensorLayout layout = ConvTensorLayout::NCHW)
{ {
using namespace ck; using namespace ck;
int N = in_nchw.mDesc.GetLengths()[0]; constexpr auto I0 = Number<0>{};
int C = in_nchw.mDesc.GetLengths()[1]; constexpr auto I1 = Number<1>{};
int HI = in_nchw.mDesc.GetLengths()[2]; constexpr auto I2 = Number<2>{};
int WI = in_nchw.mDesc.GetLengths()[3]; constexpr auto I3 = Number<3>{};
std::size_t K = wei_kcyx.mDesc.GetLengths()[0]; auto f_nchw = [&](auto n, auto c, auto hi, auto wi) {
std::size_t Y = wei_kcyx.mDesc.GetLengths()[2]; std::size_t N = in.mDesc.GetLengths()[I0];
std::size_t X = wei_kcyx.mDesc.GetLengths()[3]; std::size_t C = in.mDesc.GetLengths()[I1];
std::size_t Hi = in.mDesc.GetLengths()[I2];
std::size_t Wi = in.mDesc.GetLengths()[I3];
std::size_t HO = out_nkhw.mDesc.GetLengths()[2]; std::size_t K = wei.mDesc.GetLengths()[I0];
std::size_t WO = out_nkhw.mDesc.GetLengths()[3]; std::size_t Y = wei.mDesc.GetLengths()[I2];
std::size_t X = wei.mDesc.GetLengths()[I3];
std::size_t Ho = out.mDesc.GetLengths()[I2];
std::size_t Wo = out.mDesc.GetLengths()[I3];
auto f = [&](auto n, auto c, auto hi, auto wi) {
double v = 0; double v = 0;
for(int y = 0; y < Y; ++y) for(int y = 0; y < Y; ++y)
{ {
int h_tmp = hi + LeftPads{}[0] - y * ConvDilations{}[0]; int h_tmp = hi + in_left_pads[I0] - y * conv_dilations[I0];
if(h_tmp % ConvStrides{}[0] == 0) if(h_tmp % conv_strides[I0] == 0)
{ {
int ho = h_tmp / ConvStrides{}[0]; int ho = h_tmp / conv_strides[I0];
if(ho >= 0 && ho < HO) if(ho >= 0 && ho < Ho)
{ {
for(int x = 0; x < X; ++x) for(int x = 0; x < X; ++x)
{ {
int w_tmp = wi + LeftPads{}[1] - x * ConvDilations{}[1]; int w_tmp = wi + in_left_pads[I1] - x * conv_dilations[I1];
if(w_tmp % ConvStrides{}[1] == 0) if(w_tmp % conv_strides[I1] == 0)
{ {
int wo = w_tmp / ConvStrides{}[1]; int wo = w_tmp / conv_strides[I1];
if(wo >= 0 && wo < WO) if(wo >= 0 && wo < Wo)
{ {
for(int k = 0; k < K; ++k) for(int k = 0; k < K; ++k)
{ {
v += out_nkhw(n, k, ho, wo) * wei_kcyx(k, c, y, x); v += out(n, k, ho, wo) * wei(k, c, y, x);
} }
} }
} }
...@@ -64,14 +70,74 @@ void host_direct_convolution_backward_data(Tensor<TIn>& in_nchw, ...@@ -64,14 +70,74 @@ void host_direct_convolution_backward_data(Tensor<TIn>& in_nchw,
} }
} }
in_nchw(n, c, hi, wi) = v; in(n, c, hi, wi) = v;
}; };
auto f_par = make_ParallelTensorFunctor(f, auto f_nhwc = [&](auto n, auto hi, auto wi, auto c) {
in_nchw.mDesc.GetLengths()[0], std::size_t N = in.mDesc.GetLengths()[I0];
in_nchw.mDesc.GetLengths()[1], std::size_t Hi = in.mDesc.GetLengths()[I1];
in_nchw.mDesc.GetLengths()[2], std::size_t Wi = in.mDesc.GetLengths()[I2];
in_nchw.mDesc.GetLengths()[3]); std::size_t C = in.mDesc.GetLengths()[I3];
std::size_t K = wei.mDesc.GetLengths()[I0];
std::size_t Y = wei.mDesc.GetLengths()[I1];
std::size_t X = wei.mDesc.GetLengths()[I2];
std::size_t Ho = out.mDesc.GetLengths()[I1];
std::size_t Wo = out.mDesc.GetLengths()[I2];
double v = 0;
for(int y = 0; y < Y; ++y)
{
int h_tmp = hi + in_left_pads[I0] - y * conv_dilations[I0];
if(h_tmp % conv_strides[I0] == 0)
{
int ho = h_tmp / conv_strides[I0];
if(ho >= 0 && ho < Ho)
{
for(int x = 0; x < X; ++x)
{
int w_tmp = wi + in_left_pads[I1] - x * conv_dilations[I1];
if(w_tmp % conv_strides[I1] == 0)
{
int wo = w_tmp / conv_strides[I1];
if(wo >= 0 && wo < Wo)
{
for(int k = 0; k < K; ++k)
{
v += out(n, ho, wo, k) * wei(k, y, x, c);
}
}
}
}
}
}
}
in(n, hi, wi, c) = v;
};
f_par(std::thread::hardware_concurrency()); switch(layout)
{
case ConvTensorLayout::NCHW:
make_ParallelTensorFunctor(f_nchw,
in.mDesc.GetLengths()[0],
in.mDesc.GetLengths()[1],
in.mDesc.GetLengths()[2],
in.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
break;
case ConvTensorLayout::NHWC:
make_ParallelTensorFunctor(f_nhwc,
in.mDesc.GetLengths()[0],
in.mDesc.GetLengths()[1],
in.mDesc.GetLengths()[2],
in.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
break;
default: throw std::runtime_error("wrong! not supported layout");
}
} }
driver/include/host_tensor.hpp
View file @ 3835318c
...@@ -9,7 +9,7 @@ ...@@ -9,7 +9,7 @@
#include <cassert> #include <cassert>
#include <iostream> #include <iostream>
template <class Range> template <typename Range>
std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim) std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim)
{ {
bool first = true; bool first = true;
...@@ -24,12 +24,27 @@ std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim) ...@@ -24,12 +24,27 @@ std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim)
return os; return os;
} }
template <typename T, typename Range>
std::ostream& LogRangeAsType(std::ostream& os, Range&& range, std::string delim)
{
bool first = true;
for(auto&& v : range)
{
if(first)
first = false;
else
os << delim;
os << T{v};
}
return os;
}
typedef enum { typedef enum {
Half = 0, Half = 0,
Float = 1, Float = 1,
} DataType_t; } DataType_t;
template <class T> template <typename T>
struct DataType; struct DataType;
template <> template <>
...@@ -37,13 +52,13 @@ struct DataType<float> : std::integral_constant<DataType_t, DataType_t::Float> ...@@ -37,13 +52,13 @@ struct DataType<float> : std::integral_constant<DataType_t, DataType_t::Float>
{ {
}; };
template <class F, class T, std::size_t... Is> template <typename F, typename T, std::size_t... Is>
auto call_f_unpack_args_impl(F f, T args, std::index_sequence<Is...>) auto call_f_unpack_args_impl(F f, T args, std::index_sequence<Is...>)
{ {
return f(std::get<Is>(args)...); return f(std::get<Is>(args)...);
} }
template <class F, class T> template <typename F, typename T>
auto call_f_unpack_args(F f, T args) auto call_f_unpack_args(F f, T args)
{ {
constexpr std::size_t N = std::tuple_size<T>{}; constexpr std::size_t N = std::tuple_size<T>{};
...@@ -51,13 +66,13 @@ auto call_f_unpack_args(F f, T args) ...@@ -51,13 +66,13 @@ auto call_f_unpack_args(F f, T args)
return call_f_unpack_args_impl(f, args, std::make_index_sequence<N>{}); return call_f_unpack_args_impl(f, args, std::make_index_sequence<N>{});
} }
template <class F, class T, std::size_t... Is> template <typename F, typename T, std::size_t... Is>
auto construct_f_unpack_args_impl(T args, std::index_sequence<Is...>) auto construct_f_unpack_args_impl(T args, std::index_sequence<Is...>)
{ {
return F(std::get<Is>(args)...); return F(std::get<Is>(args)...);
} }
template <class F, class T> template <typename F, typename T>
auto construct_f_unpack_args(F, T args) auto construct_f_unpack_args(F, T args)
{ {
constexpr std::size_t N = std::tuple_size<T>{}; constexpr std::size_t N = std::tuple_size<T>{};
...@@ -77,13 +92,13 @@ struct HostTensorDescriptor ...@@ -77,13 +92,13 @@ struct HostTensorDescriptor
void CalculateStrides(); void CalculateStrides();
template <class Range> template <typename Range>
HostTensorDescriptor(const Range& lens) : mLens(lens.begin(), lens.end()) HostTensorDescriptor(const Range& lens) : mLens(lens.begin(), lens.end())
{ {
this->CalculateStrides(); this->CalculateStrides();
} }
template <class Range1, class Range2> template <typename Range1, typename Range2>
HostTensorDescriptor(const Range1& lens, const Range2& strides) HostTensorDescriptor(const Range1& lens, const Range2& strides)
: mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end()) : mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
{ {
...@@ -96,7 +111,7 @@ struct HostTensorDescriptor ...@@ -96,7 +111,7 @@ struct HostTensorDescriptor
const std::vector<std::size_t>& GetLengths() const; const std::vector<std::size_t>& GetLengths() const;
const std::vector<std::size_t>& GetStrides() const; const std::vector<std::size_t>& GetStrides() const;
template <class... Is> template <typename... Is>
std::size_t GetOffsetFromMultiIndex(Is... is) const std::size_t GetOffsetFromMultiIndex(Is... is) const
{ {
assert(sizeof...(Is) == this->GetNumOfDimension()); assert(sizeof...(Is) == this->GetNumOfDimension());
...@@ -111,7 +126,7 @@ struct HostTensorDescriptor ...@@ -111,7 +126,7 @@ struct HostTensorDescriptor
struct joinable_thread : std::thread struct joinable_thread : std::thread
{ {
template <class... Xs> template <typename... Xs>
joinable_thread(Xs&&... xs) : std::thread(std::forward<Xs>(xs)...) joinable_thread(Xs&&... xs) : std::thread(std::forward<Xs>(xs)...)
{ {
} }
...@@ -126,7 +141,7 @@ struct joinable_thread : std::thread ...@@ -126,7 +141,7 @@ struct joinable_thread : std::thread
} }
}; };
template <class F, class... Xs> template <typename F, typename... Xs>
struct ParallelTensorFunctor struct ParallelTensorFunctor
{ {
F mF; F mF;
...@@ -180,26 +195,26 @@ struct ParallelTensorFunctor ...@@ -180,26 +195,26 @@ struct ParallelTensorFunctor
} }
}; };
template <class F, class... Xs> template <typename F, typename... Xs>
auto make_ParallelTensorFunctor(F f, Xs... xs) auto make_ParallelTensorFunctor(F f, Xs... xs)
{ {
return ParallelTensorFunctor<F, Xs...>(f, xs...); return ParallelTensorFunctor<F, Xs...>(f, xs...);
} }
template <class T> template <typename T>
struct Tensor struct Tensor
{ {
template <class X> template <typename X>
Tensor(std::initializer_list<X> lens) : mDesc(lens), mData(mDesc.GetElementSpace()) Tensor(std::initializer_list<X> lens) : mDesc(lens), mData(mDesc.GetElementSpace())
{ {
} }
template <class X> template <typename X>
Tensor(std::vector<X> lens) : mDesc(lens), mData(mDesc.GetElementSpace()) Tensor(std::vector<X> lens) : mDesc(lens), mData(mDesc.GetElementSpace())
{ {
} }
template <class X, class Y> template <typename X, typename Y>
Tensor(std::vector<X> lens, std::vector<Y> strides) Tensor(std::vector<X> lens, std::vector<Y> strides)
: mDesc(lens, strides), mData(mDesc.GetElementSpace()) : mDesc(lens, strides), mData(mDesc.GetElementSpace())
{ {
...@@ -207,7 +222,7 @@ struct Tensor ...@@ -207,7 +222,7 @@ struct Tensor
Tensor(const HostTensorDescriptor& desc) : mDesc(desc), mData(mDesc.GetElementSpace()) {} Tensor(const HostTensorDescriptor& desc) : mDesc(desc), mData(mDesc.GetElementSpace()) {}
template <class G> template <typename G>
void GenerateTensorValue(G g, std::size_t num_thread = 1) void GenerateTensorValue(G g, std::size_t num_thread = 1)
{ {
switch(mDesc.GetNumOfDimension()) switch(mDesc.GetNumOfDimension())
...@@ -247,13 +262,13 @@ struct Tensor ...@@ -247,13 +262,13 @@ struct Tensor
} }
} }
template <class... Is> template <typename... Is>
T& operator()(Is... is) T& operator()(Is... is)
{ {
return mData[mDesc.GetOffsetFromMultiIndex(is...)]; return mData[mDesc.GetOffsetFromMultiIndex(is...)];
} }
template <class... Is> template <typename... Is>
const T& operator()(Is... is) const const T& operator()(Is... is) const
{ {
return mData[mDesc.GetOffsetFromMultiIndex(is...)]; return mData[mDesc.GetOffsetFromMultiIndex(is...)];
...@@ -285,7 +300,7 @@ HostTensorDescriptor::HostTensorDescriptor(std::vector<X> lens, std::vector<Y> s ...@@ -285,7 +300,7 @@ HostTensorDescriptor::HostTensorDescriptor(std::vector<X> lens, std::vector<Y> s
void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream& os = std::cout); void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream& os = std::cout);
template <class T> template <typename T>
void check_error(const Tensor<T>& ref, const Tensor<T>& result) void check_error(const Tensor<T>& ref, const Tensor<T>& result)
{ {
float error = 0; float error = 0;
......
driver/include/host_tensor_generator.hpp
View file @ 3835318c
#ifndef HOST_TENSOR_GENERATOR_HPP #ifndef HOST_TENSOR_GENERATOR_HPP
#define HOST_TENSOR_GENERATOR_HPP #define HOST_TENSOR_GENERATOR_HPP
#include <cmath>
#include "config.hpp" #include "config.hpp"
struct GeneratorTensor_1 struct GeneratorTensor_1
{ {
int value = 1; int value = 1;
template <class... Is> template <typename... Is>
double operator()(Is... is) double operator()(Is... is)
{ {
return value; return value;
...@@ -19,7 +20,7 @@ struct GeneratorTensor_2 ...@@ -19,7 +20,7 @@ struct GeneratorTensor_2
int min_value = 0; int min_value = 0;
int max_value = 1; int max_value = 1;
template <class... Is> template <typename... Is>
double operator()(Is...) double operator()(Is...)
{ {
return (std::rand() % (max_value - min_value)) + min_value; return (std::rand() % (max_value - min_value)) + min_value;
...@@ -28,7 +29,7 @@ struct GeneratorTensor_2 ...@@ -28,7 +29,7 @@ struct GeneratorTensor_2
struct GeneratorTensor_3 struct GeneratorTensor_3
{ {
template <class... Is> template <typename... Is>
double operator()(Is... is) double operator()(Is... is)
{ {
std::array<ck::index_t, sizeof...(Is)> dims = {{static_cast<ck::index_t>(is)...}}; std::array<ck::index_t, sizeof...(Is)> dims = {{static_cast<ck::index_t>(is)...}};
...@@ -41,7 +42,7 @@ struct GeneratorTensor_3 ...@@ -41,7 +42,7 @@ struct GeneratorTensor_3
struct GeneratorTensor_Checkboard struct GeneratorTensor_Checkboard
{ {
template <class... Ts> template <typename... Ts>
double operator()(Ts... Xs) const double operator()(Ts... Xs) const
{ {
std::array<ck::index_t, sizeof...(Ts)> dims = {{static_cast<ck::index_t>(Xs)...}}; std::array<ck::index_t, sizeof...(Ts)> dims = {{static_cast<ck::index_t>(Xs)...}};
......
script/cmake-rocm.sh
View file @ 3835318c
...@@ -3,24 +3,46 @@ rm -f CMakeCache.txt ...@@ -3,24 +3,46 @@ rm -f CMakeCache.txt
rm -f *.cmake rm -f *.cmake
rm -rf CMakeFiles rm -rf CMakeFiles
MY_PROJECT_SOURCE=../ MY_PROJECT_SOURCE=../../../
MY_PROJECT_INSTALL=../install.dir MY_PROJECT_INSTALL=../install.dir
cmake \ cmake \
-D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL} \ -D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL} \
-D CMAKE_BUILD_TYPE=Release \ -D CMAKE_BUILD_TYPE=Release \
-D DEVICE_BACKEND="AMD" \ -D DEVICE_BACKEND=AMD \
-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx1030 -gline-tables-only -save-temps=$CWD -ftemplate-backtrace-limit=0" \ -D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx908 -mllvm --amdgpu-spill-vgpr-to-agpr=0 -gline-tables-only -save-temps=$CWD" \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \ -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH="/opt/rocm" \ -D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \ -D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
${MY_PROJECT_SOURCE} ${MY_PROJECT_SOURCE}
#-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx1030 -gline-tables-only -save-temps=$CWD -ftemplate-backtrace-limit=0" \
#-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx1030 -gline-tables-only -save-temps=$CWD -ftemplate-backtrace-limit=0 -mllvm -print-before=amdgpu-codegenprepare -mllvm -print-module-scope" \
#-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -gline-tables-only -save-temps=$CWD" \
#-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-spill-vgpr-to-agpr=0" \ #-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-spill-vgpr-to-agpr=0" \
#-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-spill-vgpr-to-agpr=0 -save-temps=$CWD" \ #-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-spill-vgpr-to-agpr=0 -save-temps=$CWD" \
#-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-enable-global-sgpr-addr -mllvm --amdgpu-spill-vgpr-to-agpr=0" \ #-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-enable-global-sgpr-addr -mllvm --amdgpu-spill-vgpr-to-agpr=0" \
#-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-enable-global-sgpr-addr -mllvm --amdgpu-spill-vgpr-to-agpr=0 -save-temps=$CWD" \ #-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-enable-global-sgpr-addr -mllvm --amdgpu-spill-vgpr-to-agpr=0 -save-temps=$CWD" \
#-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-enable-global-sgpr-addr -mllvm --amdgpu-spill-vgpr-to-agpr=0 -v -gline-tables-only -save-temps=$CWD" \ #-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-enable-global-sgpr-addr -mllvm --amdgpu-spill-vgpr-to-agpr=0 -v -gline-tables-only -save-temps=$CWD" \
#CXX_FLAG_TMP=-Weverything
# -Wno-c++98-compat \
# -Wno-c++98-compat-pedantic \
# -Wno-conversion \
# -Wno-double-promotion \
# -Wno-exit-time-destructors \
# -Wno-extra-semi \
# -Wno-float-conversion \
# -Wno-gnu-anonymous-struct \
# -Wno-gnu-zero-variadic-macro-arguments \
# -Wno-missing-noreturn \
# -Wno-missing-prototypes \
# -Wno-nested-anon-types \
# -Wno-padded \
# -Wno-return-std-move-in-c++11 \
# -Wno-shorten-64-to-32 \
# -Wno-sign-conversion \
# -Wno-unknown-warning-option \
# -Wno-unused-command-line-argument \
# -Wno-weak-vtables \
# -Wno-covered-switch-default \
# -Wno-disabled-macro-expansion \
# -Wno-undefined-reinterpret-cast
script/cmake-rocm3.1.sh deleted 100755 → 0
View file @ 1685048a
#!/bin/bash
rm -f CMakeCache.txt
rm -f *.cmake
rm -rf CMakeFiles
MY_PROJECT_SOURCE=../../../
MY_PROJECT_INSTALL=../install.dir
cmake \
-D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL} \
-D CMAKE_BUILD_TYPE=Release \
-D DEVICE_BACKEND="AMD" \
-D CMAKE_CXX_FLAGS="--amdgpu-target=gfx906" \
-D CMAKE_CXX_COMPILER=/opt/rocm/hip/bin/hipcc \
-D CMAKE_PREFIX_PATH="/opt/rocm" \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
${MY_PROJECT_SOURCE}
#-D CMAKE_CXX_FLAGS="-gline-tables-only -v --amdgpu-target=gfx906" \
script/compile-rocm3.1.sh deleted 100755 → 0
View file @ 1685048a
#!/bin/bash
export KMOPTLLC="-mattr=+enable-ds128 -amdgpu-enable-global-sgpr-addr"
export KMDUMPISA=1
export KMDUMPLLVM=1
export KMDUMPDIR=$PWD
make -j $1
#/opt/rocm/hcc/bin/llvm-objdump -mcpu=gfx906 -source -line-numbers driver/dump-gfx906.isabin > driver/dump-gfx906.isabin.asm
script/docker-rocm3.7.sh → script/docker-rocm4.1.sh 100644 → 100755
View file @ 3835318c
...@@ -8,7 +8,7 @@ docker run \ ...@@ -8,7 +8,7 @@ docker run \
--group-add sudo \ --group-add sudo \
-w /root/workspace \ -w /root/workspace \
-v $WORKSPACE:/root/workspace \ -v $WORKSPACE:/root/workspace \
asroy/tensorflow:rocm3.7-tf2.3-dev-omp \ rocm/tensorflow:rocm4.1-tf1.15-dev \
/bin/bash /bin/bash
#--network host \ #--network host \
script/run.sh
View file @ 3835318c
#!/bin/bash #!/bin/bash
## GPU visibility
export ROCR_VISIBLE_DEVICE=0
export GPU_DEVICE_ORDINAL=0
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH ## Boost
#export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
## Compiling
export OLC_DEBUG_HIP_VERBOSE=1 export OLC_DEBUG_HIP_VERBOSE=1
export OLC_DEBUG_HIP_DUMP=1 export OLC_DEBUG_HIP_DUMP=1
export OLC_DEBUG_SAVE_TEMP_DIR=1 export OLC_DEBUG_SAVE_TEMP_DIR=1
#make -j conv_driver
#make -j conv_driver_v2 #make -j conv_driver_v2
#make -j conv_bwd_data_driver_v2
make -j conv_driver_v2_olc make -j conv_driver_v2_olc
rm -rf /root/_hip_binary_kernels_/ rm -rf /root/_hip_binary_kernels_/
...@@ -21,11 +26,21 @@ INIT=$4 ...@@ -21,11 +26,21 @@ INIT=$4
LOG=$5 LOG=$5
REPEAT=$6 REPEAT=$6
###################### layout algo verify init log repeat N__ K__ C__ Y X Hi_ Wi__ Strides Dilations LeftPads RightPads ################################ layout algo verify init log repeat N__ K___ C___ Y X Hi_ Wi__ Strides Dilations LeftPads RightPads
#driver/conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 128 192 3 3 71 71 2 2 1 1 1 1 1 1 ./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1
#driver/conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 384 192 3 3 35 35 2 2 1 1 0 0 0 0 #./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 256 1024 1 7 17 17 1 1 1 1 0 3 0 3
#driver/conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 128 128 1 7 17 17 1 1 1 1 0 3 0 3 #./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 256 256 2048 3 3 14 14 1 1 1 1 1 1 1 1
#driver/conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 256 256 3 3 14 14 1 1 1 1 1 1 1 1 #./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 256 256 256 3 3 14 14 1 1 1 1 1 1 1 1
#./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 256 512 512 3 3 7 7 1 1 1 1 1 1 1 1
#./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 128 192 3 3 71 71 2 2 1 1 1 1 1 1 #./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 512 192 3 3 35 35 2 2 1 1 0 0 0 0
./conv_driver_v2_olc $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 128 192 3 3 71 71 2 2 1 1 1 1 1 1 #./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 256 256 256 3 3 30 30 2 2 1 1 0 0 0 0
#./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 256 512 512 3 3 16 16 2 2 1 1 0 0 0 0
#./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 256 2048 1024 1 1 14 14 2 2 1 1 0 0 0 0
#./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 256 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#./conv_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 256 512 2048 1 1 7 7 1 1 1 1 0 0 0 0
#./conv_bwd_data_driver_v2 $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 256 256 256 3 3 14 14 1 1 1 1 1 1 1 1
./conv_driver_v2_olc $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1
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