Skip to content

GitLab

"...resnet50_tensorflow.git" did not exist on "4b820994191ec017f4641fde1e8fe328dc4bede4"
Unverified Commit fbdf4332 authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Add xdlops v4r4r4 into online compilation (#48) 



* init for v4r4 xdlops olc

* refactor wrap

* init impl of v4r4 nchw xdlops olc

* tuning

* test perf

* fixed v4r4 nhwc

* tuned v4r4 nhwc

* use gridwise_gemm_xdlops_v2r3

* swap a/b

* add pointer support into offline v2r3

* debugging v4r4r4 transform for olc

* change timer of olc

* refactor v4r4 xdlops nchw olc

* remove transform fun in v4r4 xdlops nhwc olc
Co-authored-by: default avatarChao Liu <chao.liu2@amd.com>
parent 0a72e4df
Expand all Show whitespace changes
Inline Side-by-side
Showing with 788 additions and 962 deletions
driver/include/device_dynamic_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk.hpp
View file @ fbdf4332
...@@ -245,6 +245,7 @@ void device_dynamic_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_k ...@@ -245,6 +245,7 @@ void device_dynamic_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_k
GemmKPerBlock, GemmKPerBlock,
GemmMPerWave, GemmMPerWave,
GemmNPerWave, GemmNPerWave,
GemmK1,
MRepeat, MRepeat,
NRepeat, NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1, GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
......
driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4_xdlops_nchw_kcyx_nkhw.hpp deleted 100644 → 0
View file @ 0a72e4df
#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
View file @ fbdf4332
...@@ -2,7 +2,7 @@ ...@@ -2,7 +2,7 @@
#include "device.hpp" #include "device.hpp"
#include "host_tensor.hpp" #include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp" #include "transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp"
#include "driver_dynamic_gemm_xdlops_v2r2.hpp" #include "driver_dynamic_gemm_xdlops_v2r3.hpp"
template <typename TInWei, template <typename TInWei,
typename TAcc, typename TAcc,
...@@ -56,63 +56,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk ...@@ -56,63 +56,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
const auto out_n_k_ho_wo_desc = const auto out_n_k_ho_wo_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(out_n_k_ho_wo_lengths); make_dynamic_naive_tensor_descriptor_packed_v2(out_n_k_ho_wo_lengths);
#if 0 #if 1
// [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 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] 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 = 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;
#elif 1
// [M, N, K0, K1] = [256, 128, 4, 8] for fp16 // [M, N, K0, K1] = [256, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256; constexpr index_t BlockSize = 256;
...@@ -120,12 +64,12 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk ...@@ -120,12 +64,12 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64; constexpr index_t GemmMPerWave = 32;
constexpr index_t GemmNPerWave = 64; constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8; constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2; constexpr index_t MRepeat = 4;
constexpr index_t NRepeat = 1; constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>; using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>; using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
...@@ -139,34 +83,6 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk ...@@ -139,34 +83,6 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1; constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8; 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 = 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 = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1; constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#endif #endif
...@@ -200,10 +116,18 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk ...@@ -200,10 +116,18 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
constexpr auto out_m0_m1_m2_n_grid_iterator_hacks = constexpr auto out_m0_m1_m2_n_grid_iterator_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, 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, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}), Sequence<0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 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, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{})); Sequence<0, 0, 2, 0, 0>{}));
...@@ -216,7 +140,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk ...@@ -216,7 +140,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
for(index_t i = 0; i < 5; ++i) for(index_t i = 0; i < 5; ++i)
{ {
float ave_time = driver_dynamic_gemm_xdlops_v2r2< float ave_time = driver_dynamic_gemm_xdlops_v2r3<
BlockSize, BlockSize,
TInWei, TInWei,
TAcc, TAcc,
...@@ -230,6 +154,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk ...@@ -230,6 +154,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
GemmKPerBlock, GemmKPerBlock,
GemmMPerWave, GemmMPerWave,
GemmNPerWave, GemmNPerWave,
GemmK1,
MRepeat, MRepeat,
NRepeat, NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1, GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
...@@ -248,15 +173,15 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk ...@@ -248,15 +173,15 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
GemmBBlockTransferSrcScalarPerVector_GemmN, GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmK1, GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy false, // don't move back src coordinate after threadwise copy
Sequence<3, 0, 1, 2>, Sequence<3, 0, 1, 2, 7, 5, 4, 6>,
3, 7,
GemmCThreadTransferDstScalarPerVector, GemmCThreadTransferDstScalarPerVector,
decltype(wei_gemmk0_gemmm_gemmk1_grid_iterator_hacks), decltype(wei_gemmk0_gemmm_gemmk1_grid_iterator_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_iterator_hacks), decltype(in_gemmk0_gemmn_gemmk1_grid_iterator_hacks),
decltype(out_m0_m1_m2_n_grid_iterator_hacks), decltype(out_m0_m1_m2_n_grid_iterator_hacks),
decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks), decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_iterator_hacks),
decltype(in_gemmk0_gemmn_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()), 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<TInWei*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()), static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
wei_gemmk0_gemmm_gemmk1_grid_desc, wei_gemmk0_gemmm_gemmk1_grid_desc,
......
driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ 0a72e4df
#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, 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 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_GemmK1 = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
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 = 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());
}
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