ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 17:00:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

adding implicit gemm v4r2

Browse Source
This commit is contained in:
Chao Liu
2019-07-05 15:35:21 -05:00
parent 923578a389
commit 8133713e96
5 changed files with 1588 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

354
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp Normal file
View File

@@ -0,0 +1,354 @@
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R1_NCHW_KCYX_NKHW
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R1_NCHW_KCYX_NKHW
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMergedTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_generic_tensor_slice_copy.hpp"
#include "blockwise_gemm.hpp"
#include "threadwise_generic_tensor_slice_copy.hpp"
namespace ck {
// define B = merge(N0, Ho, Wo)
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
index_t BPerBlock,
index_t KPerBlock,
index_t EPerBlock,
index_t N1,
index_t N2,
index_t GemmMPerThreadSubC,
index_t GemmNPerThreadSubC,
index_t GemmMLevel0Cluster,
index_t GemmNLevel0Cluster,
index_t GemmMLevel1Cluster,
index_t GemmNLevel1Cluster,
index_t GemmKPerThreadLoop,
index_t GemmDataPerReadA,
index_t GemmDataPerReadB,
class InBlockCopySubLengths_E_N1_B_N2,
class InBlockCopyClusterLengths_E_N1_B_N2,
class InBlockCopyThreadClusterArrangeOrder,
class InBlockCopySrcAccessOrder,
class InBlockCopyDstAccessOrder,
index_t InBlockCopySrcDataPerRead_B,
index_t InBlockCopyDstDataPerWrite_N2,
class WeiBlockCopySubLengths_E_K,
class WeiBlockCopyClusterLengths_E_K,
class WeiBlockCopyThreadClusterArrangeOrder,
class WeiBlockCopySrcAccessOrder,
class WeiBlockCopyDstAccessOrder,
index_t WeiBlockCopySrcDataPerRead_E,
index_t WeiBlockCopyDstDataPerWrite_K>
struct GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw
{
__device__ void Run(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global) const
{
// this is a mess
// TODO: find more elegent way of specifying (or calculating) performance parameters
static_assert(N2 == GemmNPerThreadSubC, "wrong!");
static_assert((N1 * N2 * BPerBlock) %
(GemmNPerThreadSubC * GemmNLevel0Cluster * GemmNLevel1Cluster) ==
0,
"wrong!");
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 True = integral_constant<bool, true>{};
constexpr auto in_n_c_h_w_global_desc = InGlobalDesc{};
constexpr auto wei_k_c_y_x_global_desc = WeiGlobalDesc{};
constexpr auto out_n_k_h_w_global_desc = OutGlobalDesc{};
constexpr index_t N = in_n_c_h_w_global_desc.GetLength(I0);
constexpr index_t C = in_n_c_h_w_global_desc.GetLength(I1);
constexpr index_t Hi = in_n_c_h_w_global_desc.GetLength(I2);
constexpr index_t Wi = in_n_c_h_w_global_desc.GetLength(I3);
constexpr index_t K = out_n_k_h_w_global_desc.GetLength(I1);
constexpr index_t Ho = out_n_k_h_w_global_desc.GetLength(I2);
constexpr index_t Wo = out_n_k_h_w_global_desc.GetLength(I3);
constexpr index_t Y = wei_k_c_y_x_global_desc.GetLength(I2);
constexpr index_t X = wei_k_c_y_x_global_desc.GetLength(I3);
static_assert(N % (N1 * N2) == 0, "wrong! cannot divice N evenly among thread");
constexpr index_t N0 = N / (N1 * N2);
constexpr index_t B = N0 * Ho * Wo;
constexpr index_t E = C * Y * X;
// divide block work by [K, B]
static_assert(K % KPerBlock == 0 && B % BPerBlock == 0 && E % EPerBlock == 0,
"wrong! cannot divide work evenly among block");
constexpr index_t KBlockWork = K / KPerBlock;
constexpr index_t BBlockWork = B / BPerBlock;
constexpr auto block_work_desc =
make_ConstantTensorDescriptor_packed(Sequence<KBlockWork, BBlockWork>{});
const auto block_work_multi_id =
block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
const index_t k_block_data_on_global = block_work_multi_id[0] * KPerBlock;
const index_t b_block_data_on_global = block_work_multi_id[1] * BPerBlock;
// input tensor
// tensor descriptor in device memory [N0, N1, N2, Ho, Wo]
constexpr auto in_n0_n1_n2_h_w_global_desc = in_n_c_h_w_global_desc.Slice(I2, Number<Ho>{})
.Slice(I3, Number<Wo>{})
.Fold(I0, Number<N1>{}, Number<N2>{})
.Extract(Sequence<0, 1, 2, 4, 5>{});
// batch descritpor for device memory
constexpr auto in_c_y_x_global_desc = in_n_c_h_w_global_desc.Slice(I2, Number<Y>{})
.Slice(I3, Number<X>{})
.Extract(Sequence<1, 2, 3>{});
// merged tensor descriptor in device memory [E, N1, B, N2], src of blockwise copy
constexpr auto in_e_n1_b_n2_global_merged_desc = make_ConstantMergedTensorDescriptor(
in_c_y_x_global_desc.Embed(in_n0_n1_n2_h_w_global_desc),
Sequence<0, 1, 2>{},
Sequence<4>{},
Sequence<3, 6, 7>{},
Sequence<5>{});
#if 0
if(get_block_1d_id() == 0 && get_thread_local_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_n0_n1_n2_h_w_global_desc,
"in_n0_n1_n2_h_w_global_desc: ");
print_ConstantTensorDescriptor(in_c_y_x_global_desc, "in_c_y_x_global_desc: ");
print_ConstantMergedTensorDescriptor(in_e_n1_b_n2_global_merged_desc,
"in_e_n1_b_n2_global_merged_desc: ");
}
#endif
// memory layout descriptor in LDS [E, N1, B, N2], dst of blockwise copy
// be careful of LDS alignment
constexpr auto in_e_n1_b_n2_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<EPerBlock, N1, BPerBlock, N2>{}, Number<InBlockCopyDstDataPerWrite_N2>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with multiple alignment
// requirements
static_assert(in_e_n1_b_n2_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not satisfied");
// input blockwise copy
// slice a merged tensor, reorder and copy to a normal tensor
// this copy operator already has blockwise offset built-in
auto blockwise_in_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
Float,
decltype(in_e_n1_b_n2_global_merged_desc),
decltype(in_e_n1_b_n2_block_desc),
decltype(in_e_n1_b_n2_block_desc.GetLengths()),
InBlockCopySubLengths_E_N1_B_N2,
InBlockCopyClusterLengths_E_N1_B_N2,
InBlockCopyThreadClusterArrangeOrder,
InBlockCopySrcAccessOrder,
InBlockCopyDstAccessOrder,
InBlockCopySrcDataPerRead_B,
InBlockCopyDstDataPerWrite_N2>(
{0, 0, b_block_data_on_global, 0}, {0, 0, 0, 0});
// weight tensor
// tensor descriptor in device memory, src of blockwise copy
constexpr auto wei_e_k_global_desc =
wei_k_c_y_x_global_desc.Unfold(I1, I3).ReorderGivenNew2Old(Sequence<1, 0>{});
// tensor descriptor in LDS, dst of blockwise copy
// be careful of LDS alignment
constexpr auto wei_e_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<EPerBlock, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDstDataPerWrite_K, GemmDataPerReadA)>{});
// operator for blockwise copy of weight into LDS
// slice a tensor, and copy it into another tensor
// this copy operator already have blockwise offset built-in
auto blockwise_wei_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
Float,
decltype(wei_e_k_global_desc),
decltype(wei_e_k_block_desc),
decltype(wei_e_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>(
{0, k_block_data_on_global}, {0, 0});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[EPerBlock, KPerBlock] is in LDS
// b_mtx[EPerBlocl, N1 * BPerBlock * N2] is in LDS
// c_mtx[KPerBlock, N1 * BPerBlock * N2] is distributed among threads, and saved in
// register
constexpr auto a_e_k_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<EPerBlock>{}, Number<KPerBlock>{}, Number<wei_e_k_block_desc.GetStride(I0)>{});
constexpr auto b_e_n1bn2_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<EPerBlock>{},
Number<N1 * BPerBlock * N2>{},
Number<in_e_n1_b_n2_block_desc.GetStride(I0)>{});
// sanity check
static_assert(KPerBlock % (GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster) ==
0,
"wrong!");
constexpr index_t GemmMRepeat =
KPerBlock / (GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster);
// c_thread_mtx definition: this is a mess
// TODO:: more elegent way of defining c_thread_mtx
constexpr auto c_k0k2_n1n2_thread_mtx_desc = make_ConstantMatrixDescriptor(
Number<GemmMRepeat * GemmMPerThreadSubC>{}, Number<N1 * N2>{});
const auto blockwise_gemm = BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2<
BlockSize,
decltype(a_e_k_block_mtx_desc),
decltype(b_e_n1bn2_block_mtx_desc),
decltype(c_k0k2_n1n2_thread_mtx_desc),
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB>{};
// choose GEMM implementation here
const auto run_blockwise_gemm = [&](auto... Xs) {
#if 1
return blockwise_gemm.Run(Xs...);
#else
return blockwise_gemm.Run_amd_asm(Xs...);
#endif
};
// LDS allocation for input and weight: be careful of alignment
constexpr index_t max_align = math::lcm(InBlockCopyDstDataPerWrite_N2,
WeiBlockCopyDstDataPerWrite_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr index_t in_block_space =
in_e_n1_b_n2_block_desc.GetElementSpace(Number<max_align>{});
constexpr index_t wei_block_space = wei_e_k_block_desc.GetElementSpace(Number<max_align>{});
__shared__ Float p_in_block[in_block_space];
__shared__ Float p_wei_block[wei_block_space];
// register allocation for output
Float p_out_thread[c_k0k2_n1n2_thread_mtx_desc.GetElementSpace()];
// zero out threadwise output
threadwise_matrix_set_zero(c_k0k2_n1n2_thread_mtx_desc, p_out_thread);
// do work
for(index_t e = 0; e < E; e += EPerBlock)
{
// marching slicing window
blockwise_in_copy.Run(p_in_global, p_in_block);
blockwise_wei_copy.Run(p_wei_global, p_wei_block);
__syncthreads();
run_blockwise_gemm(p_wei_block, p_in_block, p_out_thread);
__syncthreads();
blockwise_in_copy.MoveSlicingWindowOnSourceTensor(I0, Number<EPerBlock>{}, True);
blockwise_wei_copy.MoveSlicingWindowOnSourceTensor(I0, Number<EPerBlock>{}, True);
}
// copy output: register to global memory
{
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = GemmMLevel0Cluster * GemmMLevel1Cluster;
constexpr index_t K0 = K / (K1 * K2);
// define tensor descriptor for threadwise copy
// output memory layout descriptor in register
constexpr auto out_k0_k1_k2_n1_n0_h_w_n2_thread_mem_desc =
make_ConstantTensorDescriptor_packed(
Sequence<KPerBlock / (K1 * K2), 1, K2, N1, 1, 1, 1, N2>{});
// output tensor descriptor in register, src of threadwise copy
constexpr auto out_n0_n1_n2_k0_k1_k2_h_w_thread_desc =
out_k0_k1_k2_n1_n0_h_w_n2_thread_mem_desc.ReorderGivenNew2Old(
Sequence<4, 3, 7, 0, 1, 2, 5, 6>{});
// output memory layout descriptor in device memory, dst of threadwise copy
constexpr auto out_n0_n1_n2_k0_k1_k2_h_w_global_mem_desc =
out_n_k_h_w_global_desc.Fold(I1, Number<K1>{}, Number<K2>{})
.Fold(I0, Number<N1>{}, Number<N2>{});
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t k_thread_data_on_global =
k_block_data_on_global + c_thread_mtx_on_block.row;
const index_t b_thread_data_on_global =
b_block_data_on_global + c_thread_mtx_on_block.col / N2;
// output merged global tensor descriptor, for calculating origin of thread tensor
// in global memory
constexpr auto out_k_n1_b_n2_global_merged_desc = make_ConstantMergedTensorDescriptor(
out_n0_n1_n2_k0_k1_k2_h_w_global_mem_desc.Unfold(I3, I5),
Sequence<3>{},
Sequence<1>{},
Sequence<0, 4, 5>{},
Sequence<2>{});
// origin of dst in device memory
Float* p_out_thread_on_global =
p_out_global +
out_k_n1_b_n2_global_merged_desc.GetOffsetFromMultiIndex(
k_thread_data_on_global, 0, b_thread_data_on_global, 0);
threadwise_generic_tensor_slice_copy_v1(
out_n0_n1_n2_k0_k1_k2_h_w_thread_desc,
p_out_thread,
{0, 0, 0, 0, 0, 0, 0, 0},
out_n0_n1_n2_k0_k1_k2_h_w_global_mem_desc,
p_out_thread_on_global,
{0, 0, 0, 0, 0, 0, 0, 0},
out_n0_n1_n2_k0_k1_k2_h_w_thread_desc.GetLengths(),
arithmetic_sequence_gen<0, 8, 1>::type{},
Number<1>{});
}
}
};
} // namespace ck
#endif

415
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer.hpp Normal file
View File

@@ -0,0 +1,415 @@
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R1_NCHW_KCYX_NKHW_LDS_DOUBLE_BUFFER
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R1_NCHW_KCYX_NKHW_LDS_DOUBLE_BUFFER
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMergedTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_generic_tensor_slice_copy.hpp"
#include "blockwise_gemm.hpp"
#include "threadwise_generic_tensor_slice_copy.hpp"
namespace ck {
// define B = merge(N0, Ho, Wo)
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
class ConvStrides,
class ConvDilations,
index_t BPerBlock,
index_t KPerBlock,
index_t EPerBlock,
index_t N1,
index_t N2,
index_t GemmMPerThreadSubC,
index_t GemmNPerThreadSubC,
index_t GemmMLevel0Cluster,
index_t GemmNLevel0Cluster,
index_t GemmMLevel1Cluster,
index_t GemmNLevel1Cluster,
index_t GemmKPerThreadLoop,
index_t GemmDataPerReadA,
index_t GemmDataPerReadB,
class InBlockCopySubLengths_E_N1_B_N2,
class InBlockCopyClusterLengths_E_N1_B_N2,
class InBlockCopyThreadClusterArrangeOrder,
class InBlockCopySrcAccessOrder,
class InBlockCopyDstAccessOrder,
index_t InBlockCopySrcDataPerRead_B,
index_t InBlockCopyDstDataPerWrite_N2,
class WeiBlockCopySubLengths_E_K,
class WeiBlockCopyClusterLengths_E_K,
class WeiBlockCopyThreadClusterArrangeOrder,
class WeiBlockCopySrcAccessOrder,
class WeiBlockCopyDstAccessOrder,
index_t WeiBlockCopySrcDataPerRead_E,
index_t WeiBlockCopyDstDataPerWrite_K>
struct GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer
{
__device__ void Run(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global) const
{
// this is a mess
// TODO: find more elegent way of specifying (or calculating) performance parameters
static_assert(N2 == GemmNPerThreadSubC, "wrong!");
static_assert((N1 * N2 * BPerBlock) %
(GemmNPerThreadSubC * GemmNLevel0Cluster * GemmNLevel1Cluster) ==
0,
"wrong!");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I5 = Number<5>{};
constexpr auto True = integral_constant<bool, true>{};
constexpr auto in_n_c_h_w_global_desc = InGlobalDesc{};
constexpr auto wei_k_c_y_x_global_desc = WeiGlobalDesc{};
constexpr auto out_n_k_h_w_global_desc = OutGlobalDesc{};
constexpr index_t N = in_n_c_h_w_global_desc.GetLength(I0);
constexpr index_t C = in_n_c_h_w_global_desc.GetLength(I1);
constexpr index_t K = out_n_k_h_w_global_desc.GetLength(I1);
constexpr index_t Ho = out_n_k_h_w_global_desc.GetLength(I2);
constexpr index_t Wo = out_n_k_h_w_global_desc.GetLength(I3);
constexpr index_t Y = wei_k_c_y_x_global_desc.GetLength(I2);
constexpr index_t X = wei_k_c_y_x_global_desc.GetLength(I3);
constexpr index_t ConvStrideH = ConvStrides{}[0];
constexpr index_t ConvStrideW = ConvStrides{}[1];
constexpr index_t ConvDilationH = ConvDilations{}[0];
constexpr index_t ConvDilationW = ConvDilations{}[1];
static_assert(N % (N1 * N2) == 0, "wrong! cannot divice N evenly among thread");
constexpr index_t N0 = N / (N1 * N2);
constexpr index_t B = N0 * Ho * Wo;
constexpr index_t E = C * Y * X;
// sanity-check for vectorized memory load
static_assert(ConvStrideW == 1 || InBlockCopySrcDataPerRead_B == 1,
"wrong! global vector load of input tensor is wrong");
static_assert((X == 1 || ConvDilationW % InBlockCopySrcDataPerRead_B == 0),
"wrong! aligment requirement for vectorized global load of input tensor will "
"be violated");
// divide block work by [K, B]
static_assert(K % KPerBlock == 0 && B % BPerBlock == 0 && E % (2 * EPerBlock) == 0,
"wrong! cannot divide work evenly among block");
constexpr index_t KBlockWork = K / KPerBlock;
constexpr index_t BBlockWork = B / BPerBlock;
constexpr auto block_work_desc =
make_ConstantTensorDescriptor_packed(Sequence<KBlockWork, BBlockWork>{});
const auto block_work_multi_id =
block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
const index_t k_block_data_on_global = block_work_multi_id[0] * KPerBlock;
const index_t b_block_data_on_global = block_work_multi_id[1] * BPerBlock;
// input tensor
// tensor descriptor in device memory [N0, N1, N2, Ho, Wo]
constexpr auto in_n0_n1_n2_h_w_global_desc =
in_n_c_h_w_global_desc.StridedSlice(I2, Number<Ho>{}, Number<ConvStrideH>{})
.StridedSlice(I3, Number<Wo>{}, Number<ConvStrideW>{})
.Fold(I0, Number<N1>{}, Number<N2>{})
.Extract(Sequence<0, 1, 2, 4, 5>{});
// batch descritpor for device memory
constexpr auto in_c_y_x_global_desc =
in_n_c_h_w_global_desc.StridedSlice(I2, Number<Y>{}, Number<ConvDilationH>{})
.StridedSlice(I3, Number<X>{}, Number<ConvDilationW>{})
.Extract(Sequence<1, 2, 3>{});
// merged tensor descriptor in device memory [E, N1, B, N2], src of blockwise copy
constexpr auto in_e_n1_b_n2_global_merged_desc = make_ConstantMergedTensorDescriptor(
in_c_y_x_global_desc.Embed(in_n0_n1_n2_h_w_global_desc),
Sequence<0, 1, 2>{},
Sequence<4>{},
Sequence<3, 6, 7>{},
Sequence<5>{});
// memory layout descriptor in LDS [E, N1, B, N2], dst of blockwise copy
// be careful of LDS alignment
constexpr auto in_e_n1_b_n2_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<EPerBlock, N1, BPerBlock, N2>{}, Number<InBlockCopyDstDataPerWrite_N2>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with multiple alignment
// requirements
static_assert(in_e_n1_b_n2_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not satisfied");
// input blockwise copy
// slice a merged tensor, reorder and copy to a normal tensor
// this copy operator already has blockwise offset built-in
auto blockwise_in_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
Float,
decltype(in_e_n1_b_n2_global_merged_desc),
decltype(in_e_n1_b_n2_block_desc),
decltype(in_e_n1_b_n2_block_desc.GetLengths()),
InBlockCopySubLengths_E_N1_B_N2,
InBlockCopyClusterLengths_E_N1_B_N2,
InBlockCopyThreadClusterArrangeOrder,
InBlockCopySrcAccessOrder,
InBlockCopyDstAccessOrder,
InBlockCopySrcDataPerRead_B,
InBlockCopyDstDataPerWrite_N2>(
{0, 0, b_block_data_on_global, 0}, {0, 0, 0, 0});
// weight tensor
// tensor descriptor in device memory, src of blockwise copy
constexpr auto wei_e_k_global_desc =
wei_k_c_y_x_global_desc.Unfold(I1, I3).ReorderGivenNew2Old(Sequence<1, 0>{});
// tensor descriptor in LDS, dst of blockwise copy
// be careful of LDS alignment
constexpr auto wei_e_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<EPerBlock, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDstDataPerWrite_K, GemmDataPerReadA)>{});
// operator for blockwise copy of weight into LDS
// slice a tensor, and copy it into another tensor
// this copy operator already have blockwise offset built-in
auto blockwise_wei_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
Float,
decltype(wei_e_k_global_desc),
decltype(wei_e_k_block_desc),
decltype(wei_e_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>(
{0, k_block_data_on_global}, {0, 0});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[EPerBlock, KPerBlock] is in LDS
// b_mtx[EPerBlocl, N1 * BPerBlock * N2] is in LDS
// c_mtx[KPerBlock, N1 * BPerBlock * N2] is distributed among threads, and saved in
// register
constexpr auto a_e_k_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<EPerBlock>{}, Number<KPerBlock>{}, Number<wei_e_k_block_desc.GetStride(I0)>{});
constexpr auto b_e_n1bn2_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<EPerBlock>{},
Number<N1 * BPerBlock * N2>{},
Number<in_e_n1_b_n2_block_desc.GetStride(I0)>{});
// sanity check
static_assert(KPerBlock % (GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster) ==
0,
"wrong!");
constexpr index_t GemmMRepeat =
KPerBlock / (GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster);
// c_thread_mtx definition: this is a mess
// TODO:: more elegent way of defining c_thread_mtx
constexpr auto c_k0k2_n1n2_thread_mtx_desc = make_ConstantMatrixDescriptor(
Number<GemmMRepeat * GemmMPerThreadSubC>{}, Number<N1 * N2>{});
const auto blockwise_gemm = BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2<
BlockSize,
decltype(a_e_k_block_mtx_desc),
decltype(b_e_n1bn2_block_mtx_desc),
decltype(c_k0k2_n1n2_thread_mtx_desc),
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB>{};
// LDS allocation for input and weight: be careful of alignment
constexpr index_t max_align = math::lcm(InBlockCopyDstDataPerWrite_N2,
WeiBlockCopyDstDataPerWrite_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr index_t in_block_space =
math::integer_least_multiple(in_e_n1_b_n2_block_desc.GetElementSpace(), max_align);
constexpr index_t wei_block_space =
math::integer_least_multiple(wei_e_k_block_desc.GetElementSpace(), max_align);
__shared__ Float p_in_block_double[2 * in_block_space];
__shared__ Float p_wei_block_double[2 * wei_block_space];
// register allocation for output
Float p_out_thread[c_k0k2_n1n2_thread_mtx_desc.GetElementSpace()];
// zero out threadwise output
threadwise_matrix_set_zero(c_k0k2_n1n2_thread_mtx_desc, p_out_thread);
const Float* p_wei_block_on_global = p_wei_global;
// LDS double buffer: preload data into LDS
{
blockwise_in_copy.Run(p_in_global, p_in_block_double);
blockwise_wei_copy.Run(p_wei_global, p_wei_block_double);
}
// LDS double buffer: main body
for(index_t e_block_data_begin = 0; e_block_data_begin + 2 * EPerBlock < E;
e_block_data_begin += 2 * EPerBlock)
{
#pragma unroll
for(index_t iloop = 0; iloop < 2; ++iloop)
{
const bool even_loop = (iloop % 2 == 0);
Float* p_in_block_now =
even_loop ? p_in_block_double : p_in_block_double + in_block_space;
Float* p_wei_block_now =
even_loop ? p_wei_block_double : p_wei_block_double + wei_block_space;
Float* p_in_block_next =
even_loop ? p_in_block_double + in_block_space : p_in_block_double;
Float* p_wei_block_next =
even_loop ? p_wei_block_double + wei_block_space : p_wei_block_double;
Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
blockwise_in_copy.MoveSlicingWindowOnSourceTensor(I0, Number<EPerBlock>{}, True);
p_wei_block_on_global += EPerBlock * wei_e_k_global_desc.GetStride(I0);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global, p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_block_on_global,
p_wei_register_clipboard);
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(p_wei_block_now, p_in_block_now, p_out_thread);
// LDS double buffer: store next data to LDS
blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard,
p_in_block_next);
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard,
p_wei_block_next);
}
}
// LDS double buffer: tail
{
Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
// even iteration
blockwise_in_copy.MoveSlicingWindowOnSourceTensor(I0, Number<EPerBlock>{}, True);
p_wei_block_on_global += EPerBlock * wei_e_k_global_desc.GetStride(I0);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global, p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_block_on_global,
p_wei_register_clipboard);
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(p_wei_block_double, p_in_block_double, p_out_thread);
// LDS double buffer: store next data to LDS
blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard,
p_in_block_double + in_block_space);
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard,
p_wei_block_double + wei_block_space);
// odd iteration
__syncthreads();
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(p_wei_block_double + wei_block_space,
p_in_block_double + in_block_space,
p_out_thread);
}
// copy output: register to global memory
{
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = GemmMLevel0Cluster * GemmMLevel1Cluster;
// define tensor descriptor for threadwise copy
// output memory layout descriptor in register
constexpr auto out_k0_k1_k2_n1_n0_h_w_n2_thread_mem_desc =
make_ConstantTensorDescriptor_packed(
Sequence<KPerBlock / (K1 * K2), 1, K2, N1, 1, 1, 1, N2>{});
// output tensor descriptor in register, src of threadwise copy
constexpr auto out_n0_n1_n2_k0_k1_k2_h_w_thread_desc =
out_k0_k1_k2_n1_n0_h_w_n2_thread_mem_desc.ReorderGivenNew2Old(
Sequence<4, 3, 7, 0, 1, 2, 5, 6>{});
// output memory layout descriptor in device memory, dst of threadwise copy
constexpr auto out_n0_n1_n2_k0_k1_k2_h_w_global_mem_desc =
out_n_k_h_w_global_desc.Fold(I1, Number<K1>{}, Number<K2>{})
.Fold(I0, Number<N1>{}, Number<N2>{});
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t k_thread_data_on_global =
k_block_data_on_global + c_thread_mtx_on_block.row;
const index_t b_thread_data_on_global =
b_block_data_on_global + c_thread_mtx_on_block.col / N2;
// output merged global tensor descriptor, for calculating origin of thread tensor
// in global memory
constexpr auto out_k_n1_b_n2_global_merged_desc = make_ConstantMergedTensorDescriptor(
out_n0_n1_n2_k0_k1_k2_h_w_global_mem_desc.Unfold(I3, I5),
Sequence<3>{},
Sequence<1>{},
Sequence<0, 4, 5>{},
Sequence<2>{});
// origin of dst in device memory
Float* p_out_thread_on_global =
p_out_global +
out_k_n1_b_n2_global_merged_desc.GetOffsetFromMultiIndex(
k_thread_data_on_global, 0, b_thread_data_on_global, 0);
threadwise_generic_tensor_slice_copy_v1(
out_n0_n1_n2_k0_k1_k2_h_w_thread_desc,
p_out_thread,
{0, 0, 0, 0, 0, 0, 0, 0},
out_n0_n1_n2_k0_k1_k2_h_w_global_mem_desc,
p_out_thread_on_global,
{0, 0, 0, 0, 0, 0, 0, 0},
out_n0_n1_n2_k0_k1_k2_h_w_thread_desc.GetLengths(),
arithmetic_sequence_gen<0, 8, 1>::type{},
Number<1>{});
}
}
};
} // namespace ck
#endif

428
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw_lds_double_buffer.hpp Normal file
View File

@@ -0,0 +1,428 @@
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R2_NCHW_KCYX_NKHW_LDS_DOUBLE_BUFFER
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R2_NCHW_KCYX_NKHW_LDS_DOUBLE_BUFFER
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMergedTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_generic_tensor_slice_copy.hpp"
#include "blockwise_gemm.hpp"
#include "threadwise_generic_tensor_slice_copy.hpp"
namespace ck {
// define B = merge(N0, Ho, Wo)
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
class ConvStrides,
class ConvDilations,
index_t BPerBlock,
index_t KPerBlock,
index_t EPerBlock,
index_t N0,
index_t N2,
index_t Ho0,
index_t Ho2,
index_t Wo0,
index_t Wo2,
index_t GemmMPerThreadSubC,
index_t GemmNPerThreadSubC,
index_t GemmMLevel0Cluster,
index_t GemmNLevel0Cluster,
index_t GemmMLevel1Cluster,
index_t GemmNLevel1Cluster,
index_t GemmKPerThreadLoop,
index_t GemmDataPerReadA,
index_t GemmDataPerReadB,
class InBlockCopySubLengths_E_N1_B_N2,
class InBlockCopyClusterLengths_E_N1_B_N2,
class InBlockCopyThreadClusterArrangeOrder,
class InBlockCopySrcAccessOrder,
class InBlockCopyDstAccessOrder,
index_t InBlockCopySrcDataPerRead_B,
index_t InBlockCopyDstDataPerWrite_N2,
class WeiBlockCopySubLengths_E_K,
class WeiBlockCopyClusterLengths_E_K,
class WeiBlockCopyThreadClusterArrangeOrder,
class WeiBlockCopySrcAccessOrder,
class WeiBlockCopyDstAccessOrder,
index_t WeiBlockCopySrcDataPerRead_E,
index_t WeiBlockCopyDstDataPerWrite_K>
struct GridwiseConvolutionImplicitGemm_v4r2_nchw_kcyx_nkhw_lds_double_buffer
{
__device__ void Run(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global) const
{
// this is a mess
// TODO: find more elegent way of specifying (or calculating) performance parameters
static_assert(N2 == GemmNPerThreadSubC, "wrong!");
static_assert((N1 * N2 * BPerBlock) %
(GemmNPerThreadSubC * GemmNLevel0Cluster * GemmNLevel1Cluster) ==
0,
"wrong!");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I5 = Number<5>{};
constexpr auto True = integral_constant<bool, true>{};
constexpr auto in_n_c_h_w_global_desc = InGlobalDesc{};
constexpr auto wei_k_c_y_x_global_desc = WeiGlobalDesc{};
constexpr auto out_n_k_h_w_global_desc = OutGlobalDesc{};
constexpr index_t N = in_n_c_h_w_global_desc.GetLengths()[0];
constexpr index_t C = in_n_c_h_w_global_desc.GetLengths()[1];
constexpr index_t K = out_n_k_h_w_global_desc.GetLengths()[1];
constexpr index_t Ho = out_n_k_h_w_global_desc.GetLengths()[2];
constexpr index_t Wo = out_n_k_h_w_global_desc.GetLengths()[3];
constexpr index_t Y = wei_k_c_y_x_global_desc.GetLengths()[2];
constexpr index_t X = wei_k_c_y_x_global_desc.GetLengths()[3];
constexpr index_t ConvStrideH = ConvStrides{}[0];
constexpr index_t ConvStrideW = ConvStrides{}[1];
constexpr index_t ConvDilationH = ConvDilations{}[0];
constexpr index_t ConvDilationW = ConvDilations{}[1];
constexpr index_t E = C * Y * X;
constexpr index_t N1 = N / (N0 * N2);
constexpr index_t Ho1 = Ho / (Ho0 * Ho2);
constexpr index_t Wo1 = Wo / (Wo0 * Wo2);
constexpr index_t B1 = N1 * Ho1 * Wo1;
static_assert((X == 1 || ConvDilationW % InBlockCopySrcDataPerRead_B == 0),
"wrong! aligment requirement for vectorized global load of input tensor will "
"be violated");
// divide block work by [K, B]
static_assert(K % KPerBlock == 0 && B % BPerBlock == 0 && E % (2 * EPerBlock) == 0,
"wrong! cannot divide work evenly among block");
constexpr index_t KBlockWork = K / KPerBlock;
constexpr index_t B1BlockWork = B1 / B1PerBlock;
constexpr auto block_work_desc =
make_ConstantTensorDescriptor_packed(Sequence<KBlockWork, B1BlockWork>{});
const auto block_work_multi_id =
block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
const index_t k_block_data_on_global = block_work_multi_id[0] * KPerBlock;
const index_t b1_block_data_on_global = block_work_multi_id[1] * B1PerBlock;
// input tensor
// tensor descriptor in device memory [N0, N1, N2, Ho0, Ho1, Ho2, Wo0, Wo1, Wo2]
constexpr auto in_n0_n1_n2_ho0_ho1_ho2_wo0_wo1_wo2_global_desc =
in_n_c_h_w_global_desc.Extract(I0, I2, I3)
.StridedSlice(I1, Number<Ho>{}, Number<ConvStrideH>{})
.StridedSlice(I2, Number<Wo>{}, Number<ConvStrideW>{})
.Fold(I2, Number<Wo1>{}, Number<Wo2>{})
.Fold(I1, Number<Ho1>{}, Number<Ho2>{})
.Fold(I0, Number<N1>{}, Number<N2>{});
constexpr auto in_n0_ho0_wo0_n1_ho1_wo1_n2_ho2_wo2_global_desc =
in_n0_n1_n2_ho0_ho1_ho2_wo0_wo1_wo2_global_desc.ReorderGivenNew2Old(
Sequence<0, 3, 6, 1, 4, 7, 2, 5, 8>{});
// batch descritpor for device memory
constexpr auto in_c_y_x_global_desc =
in_n_c_h_w_global_desc.StridedSlice(I2, Number<Y>{}, Number<ConvDilationH>{})
.StridedSlice(I3, Number<X>{}, Number<ConvDilationW>{})
.Extract(Sequence<1, 2, 3>{});
// merged tensor descriptor in device memory [E, N1, B, N2], src of blockwise copy
constexpr auto in_e_n0_ho0_wo0_b1_n2_ho2_wo2_global_merged_desc =
make_ConstantMergedTensorDescriptor(
in_c_y_x_global_desc.Embed(in_n0_ho0_wo0_n1_ho1_wo1_n2_ho2_wo2_global_desc),
Sequence<0, 1, 2>{},
Sequence<3>{},
Sequence<4>{},
Sequence<5>{},
Sequence<6, 7, 8>{},
Sequence<9>{},
Sequence<10>{},
Sequence<11>{});
// memory layout descriptor in LDS [E, N1, B, N2], dst of blockwise copy
// be careful of LDS alignment
constexpr auto in_e_n0_ho0_wo0_b1_n2_ho2_wo2_block_desc =
in_e_n0_ho0_wo0_b1_n2_ho2_wo2_global_merged_desc.Pack();
// input blockwise copy
// slice a merged tensor, reorder and copy to a normal tensor
// this copy operator already has blockwise offset built-in
auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1<
BlockSize,
Float,
decltype(in_e_n0_ho0_wo0_b1_n2_ho2_wo2_global_merged_desc),
decltype(in_e_n0_ho0_wo0_b1_n2_ho2_wo2_block_desc),
decltype(in_e_n0_ho0_wo0_b1_n2_ho2_wo2_block_desc.GetLengths()),
InBlockCopySubLengths_E_N0_Ho0_Wo0_B1_N2_Ho2_Wo2,
InBlockCopyClusterLengths_E_N0_Ho0_Wo0_B1_N2_Ho2_Wo2,
InBlockCopyThreadClusterArrangeOrder,
InBlockCopySrcAccessOrder,
InBlockCopyDstAccessOrder,
InBlockCopyDataPerAccess_Wo2,
InBlockCopyDataPerAccess_Wo2>({0, 0, 0, 0, b1_block_data_on_global, 0, 0, 0},
{0, 0, 0, 0, 0, 0, 0, 0});
// weight tensor
// tensor descriptor in device memory, src of blockwise copy
constexpr auto wei_e_k_global_desc =
wei_k_c_y_x_global_desc.Unfold(I1, I3).ReorderGivenNew2Old(Sequence<1, 0>{});
// tensor descriptor in LDS, dst of blockwise copy
// be careful of LDS alignment
constexpr auto wei_e_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<EPerBlock, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDstDataPerWrite_K, GemmDataPerReadA)>{});
// operator for blockwise copy of weight into LDS
// slice a tensor, and copy it into another tensor
// this copy operator already have blockwise offset built-in
auto blockwise_wei_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
Float,
decltype(wei_e_k_global_desc),
decltype(wei_e_k_block_desc),
decltype(wei_e_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>(
{0, k_block_data_on_global}, {0, 0});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[EPerBlock, KPerBlock] is in LDS
// b_mtx[EPerBlocl, N1 * BPerBlock * N2] is in LDS
// c_mtx[KPerBlock, N1 * BPerBlock * N2] is distributed among threads, and saved in
// register
constexpr auto a_e_k_block_mtx_desc =
make_ConstantMatrixDescriptor_from_ConstantTensorDescriptor(wei_e_k_block_desc);
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with multiple alignment
// requirements
static_assert(in_e_n0_ho0_wo0_b1_n2_ho2_wo2_block_desc.GetStrides()[3] % GemmDataPerReadB ==
0,
"GemmDataPerReadB alignment requirement is not satisfied");
constexpr auto b_e_n0ho0wo0b1n2ho2wo2_block_mtx_desc =
make_ConstantMatrixDescriptor_from_ConstantTensorDescriptor(
in_e_n0_ho0_wo0_b1_n2_ho2_wo2_block_desc.Unfold(I1, I7));
// sanity check
static_assert(KPerBlock % (GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster) ==
0,
"wrong!");
constexpr index_t GemmMRepeat =
KPerBlock / (GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster);
// c_thread_mtx definition: this is a mess
// TODO:: more elegent way of defining c_thread_mtx
constexpr auto c_k0k2_n0ho0wo0n2ho2wo2_thread_mtx_desc = make_ConstantMatrixDescriptor(
Number<GemmMRepeat * GemmMPerThreadSubC>{}, Number<N0 * Ho0 * Wo0 * N2 * Ho2 * Wo2>{});
const auto blockwise_gemm = BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2<
BlockSize,
decltype(a_e_k_block_mtx_desc),
decltype(b_e_n0ho0wo0b1n2ho2wo2_block_mtx_desc),
decltype(c_k0k2_n0ho0wo0n2ho2wo2_thread_mtx_desc),
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB>{};
// LDS allocation for input and weight: be careful of alignment
constexpr index_t max_align = math::lcm(InBlockCopyDstDataPerWrite_N2,
WeiBlockCopyDstDataPerWrite_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr index_t in_block_space = math::integer_least_multiple(
in_e_n0_ho0_wo0_b1_n2_ho2_wo2_block_desc.GetElementSpace(), max_align);
constexpr index_t wei_block_space =
math::integer_least_multiple(wei_e_k_block_desc.GetElementSpace(), max_align);
__shared__ Float p_in_block_double[2 * in_block_space];
__shared__ Float p_wei_block_double[2 * wei_block_space];
// register allocation for output
Float p_out_thread[c_k0k2_n0ho0wo0n2ho2wo2_thread_mtx_desc.GetElementSpace()];
// zero out threadwise output
threadwise_matrix_set_zero(c_k0k2_n0ho0wo0n2ho2wo2_thread_mtx_desc, p_out_thread);
const Float* p_wei_block_on_global = p_wei_global;
// LDS double buffer: preload data into LDS
{
blockwise_in_copy.Run(p_in_global, p_in_block_double);
blockwise_wei_copy.Run(p_wei_global, p_wei_block_double);
}
// LDS double buffer: main body
for(index_t e_block_data_begin = 0; e_block_data_begin + 2 * EPerBlock < E;
e_block_data_begin += 2 * EPerBlock)
{
#pragma unroll
for(index_t iloop = 0; iloop < 2; ++iloop)
{
const bool even_loop = (iloop % 2 == 0);
Float* p_in_block_now =
even_loop ? p_in_block_double : p_in_block_double + in_block_space;
Float* p_wei_block_now =
even_loop ? p_wei_block_double : p_wei_block_double + wei_block_space;
Float* p_in_block_next =
even_loop ? p_in_block_double + in_block_space : p_in_block_double;
Float* p_wei_block_next =
even_loop ? p_wei_block_double + wei_block_space : p_wei_block_double;
Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
blockwise_in_copy.MoveSlicingWindowOnSourceTensor(I0, Number<EPerBlock>{}, True);
p_wei_block_on_global += EPerBlock * wei_e_k_global_desc.GetStride(I0);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global, p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_block_on_global,
p_wei_register_clipboard);
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(p_wei_block_now, p_in_block_now, p_out_thread);
// LDS double buffer: store next data to LDS
blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard,
p_in_block_next);
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard,
p_wei_block_next);
}
}
// LDS double buffer: tail
{
Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
// even iteration
blockwise_in_copy.MoveSlicingWindowOnSourceTensor(I0, Number<EPerBlock>{}, True);
p_wei_block_on_global += EPerBlock * wei_e_k_global_desc.GetStride(I0);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global, p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_block_on_global,
p_wei_register_clipboard);
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(p_wei_block_double, p_in_block_double, p_out_thread);
// LDS double buffer: store next data to LDS
blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard,
p_in_block_double + in_block_space);
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard,
p_wei_block_double + wei_block_space);
// odd iteration
__syncthreads();
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(p_wei_block_double + wei_block_space,
p_in_block_double + in_block_space,
p_out_thread);
}
// copy output: register to global memory
{
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = GemmMLevel0Cluster * GemmMLevel1Cluster;
// define tensor descriptor for threadwise copy
// output memory layout descriptor in register
constexpr auto out_k0_k1_k2_n0_ho0_wo0_n1_ho1_wo1_n2_ho2_wo2_thread_mem_desc =
make_ConstantTensorDescriptor_packed(
Sequence<KPerBlock / (K1 * K2), 1, K2, N0, ho0, wo0, 1, 1, 1, N2, Ho2, Wo2>{});
// output tensor descriptor in register, src of threadwise copy
constexpr auto out_n0_n1_n2_k0_k1_k2_ho0_ho1_ho2_wo0_wo1_wo2_thread_desc =
out_k0_k1_k2_n0_ho0_wo0_n1_ho1_wo1_n2_ho2_wo2_thread_mem_desc.ReorderGivenNew2Old(
Sequence<3, 6, 9, 0, 1, 2, 4, 7, 10, 5, 8, 11>{});
// output memory layout descriptor in device memory, dst of threadwise copy
constexpr auto out_n0_n1_n2_k0_k1_k2_ho0_ho1_ho2_wo0_wo1_wo2_global_desc =
out_n_k_h_w_global_desc.Fold(I3, Sequence<Wo1, Wo2>{})
.Fold(I2, Sequence<Ho1, Ho2>{})
.Fold(I1, Sequence<K1, K2>{})
.Fold(I0, Sequence<N1, N2>{});
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t k_thread_data_on_global =
k_block_data_on_global + c_thread_mtx_on_block.row;
const index_t b_thread_data_on_global =
b_block_data_on_global + c_thread_mtx_on_block.col / N2;
// output merged global tensor descriptor, for calculating origin of thread tensor
// in global memory
constexpr auto out_k_n1_b_n2_global_merged_desc = make_ConstantMergedTensorDescriptor(
out_n0_n1_n2_k0_k1_k2_h_w_global_mem_desc.Unfold(I3, I5),
Sequence<3>{},
Sequence<1>{},
Sequence<0, 4, 5>{},
Sequence<2>{});
// origin of dst in device memory
Float* p_out_thread_on_global =
p_out_global +
out_k_n1_b_n2_global_merged_desc.GetOffsetFromMultiIndex(
k_thread_data_on_global, 0, b_thread_data_on_global, 0);
threadwise_generic_tensor_slice_copy_v1(
out_n0_n1_n2_k0_k1_k2_h_w_thread_desc,
p_out_thread,
{0, 0, 0, 0, 0, 0, 0, 0},
out_n0_n1_n2_k0_k1_k2_h_w_global_mem_desc,
p_out_thread_on_global,
{0, 0, 0, 0, 0, 0, 0, 0},
out_n0_n1_n2_k0_k1_k2_h_w_thread_desc.GetLengths(),
arithmetic_sequence_gen<0, 8, 1>::type{},
Number<1>{});
}
}
};
} // namespace ck
#endif

232
driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp Normal file
View File

@@ -0,0 +1,232 @@
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "tensor.hpp"
#include "gridwise_convolution_kernel_wrapper.hpp"
#include "gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp"
#include "gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer.hpp"
using namespace ck;
template <class T,
class InDesc,
class WeiDesc,
class OutDesc,
class ConvStrides,
class ConvDilations>
void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
const Tensor<T>& in_nchw,
WeiDesc,
const Tensor<T>& wei_kcyx,
OutDesc,
Tensor<T>& out_nkhw,
ConvStrides,
ConvDilations,
index_t nrepeat)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_nchw_desc = InDesc{};
constexpr auto wei_kcyx_desc = WeiDesc{};
constexpr auto out_nkhw_desc = OutDesc{};
constexpr index_t Hi = in_nchw_desc.GetLength(I2);
constexpr index_t Wi = in_nchw_desc.GetLength(I3);
constexpr index_t N = out_nkhw_desc.GetLength(I0);
constexpr index_t Ho = out_nkhw_desc.GetLength(I2);
constexpr index_t Wo = out_nkhw_desc.GetLength(I3);
constexpr index_t K = wei_kcyx_desc.GetLength(I0);
constexpr index_t C = wei_kcyx_desc.GetLength(I1);
constexpr index_t Y = wei_kcyx_desc.GetLength(I2);
constexpr index_t X = wei_kcyx_desc.GetLength(I3);
std::size_t data_sz = sizeof(T);
DeviceMem in_nchw_device_buf(data_sz * in_nchw.mDesc.GetElementSpace());
DeviceMem wei_kcyx_device_buf(data_sz * wei_kcyx.mDesc.GetElementSpace());
DeviceMem out_nkhw_device_buf(data_sz * out_nkhw.mDesc.GetElementSpace());
in_nchw_device_buf.ToDevice(in_nchw.mData.data());
wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
constexpr index_t N1 = 2;
constexpr index_t N2 = 4;
constexpr index_t B = (N * Ho * Wo) / (N1 * N2);
#if 1
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 16;
constexpr index_t KPerBlock = 128;
constexpr index_t EPerBlock = 8;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 4;
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopySubLengths_E_N1_B_N2 = Sequence<1, 1, 1, 4>;
using InBlockCopyClusterLengths_E_N1_B_N2 = Sequence<8, 2, 16, 1>;
using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopySrcAccessOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopyDstAccessOrder = Sequence<0, 1, 2, 3>; // [E, N1, B, N2]
constexpr index_t InBlockCopySrcDataPerRead_B = 1;
constexpr index_t InBlockCopyDstDataPerWrite_N2 = 4;
using WeiBlockCopySubLengths_E_K = Sequence<4, 1>;
using WeiBlockCopyClusterLengths_E_K = Sequence<2, 128>;
using WeiBlockCopyThreadClusterArrangeOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopySrcAccessOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopyDstAccessOrder = Sequence<0, 1>; // [E, K]
constexpr index_t WeiBlockCopySrcDataPerRead_E = 4;
constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
#elif 0
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 16;
constexpr index_t KPerBlock = 128;
constexpr index_t EPerBlock = 8;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 4;
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopySubLengths_E_N1_B_N2 = Sequence<1, 1, 4, 1>;
using InBlockCopyClusterLengths_E_N1_B_N2 = Sequence<8, 2, 4, 4>;
using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopySrcAccessOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopyDstAccessOrder = Sequence<0, 1, 2, 3>; // [E, N1, B, N2]
constexpr index_t InBlockCopySrcDataPerRead_B = 4;
constexpr index_t InBlockCopyDstDataPerWrite_N2 = 1;
using WeiBlockCopySubLengths_E_K = Sequence<4, 1>;
using WeiBlockCopyClusterLengths_E_K = Sequence<2, 128>;
using WeiBlockCopyThreadClusterArrangeOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopySrcAccessOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopyDstAccessOrder = Sequence<0, 1>; // [E, K]
constexpr index_t WeiBlockCopySrcDataPerRead_E = 4;
constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
#elif 1
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 16;
constexpr index_t KPerBlock = 128;
constexpr index_t EPerBlock = 8;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 4;
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopySubLengths_E_N1_B_N2 = Sequence<1, 1, 2, 2>;
using InBlockCopyClusterLengths_E_N1_B_N2 = Sequence<8, 2, 8, 2>;
using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopySrcAccessOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopyDstAccessOrder = Sequence<0, 1, 2, 3>; // [E, N1, B, N2]
constexpr index_t InBlockCopySrcDataPerRead_B = 2;
constexpr index_t InBlockCopyDstDataPerWrite_N2 = 2;
using WeiBlockCopySubLengths_E_K = Sequence<4, 1>;
using WeiBlockCopyClusterLengths_E_K = Sequence<2, 128>;
using WeiBlockCopyThreadClusterArrangeOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopySrcAccessOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopyDstAccessOrder = Sequence<0, 1>; // [E, K]
constexpr index_t WeiBlockCopySrcDataPerRead_E = 4;
constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
#endif
constexpr index_t GridSize =
((B + BPerBlock - 1) / BPerBlock) * ((K + KPerBlock - 1) / KPerBlock);
printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
for(index_t i = 0; i < nrepeat; ++i)
{
constexpr auto gridwise_conv =
#if 0
GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw
#else
GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer
#endif
<GridSize,
BlockSize,
T,
decltype(in_nchw_desc),
decltype(wei_kcyx_desc),
decltype(out_nkhw_desc),
ConvStrides,
ConvDilations,
BPerBlock,
KPerBlock,
EPerBlock,
N1,
N2,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB,
InBlockCopySubLengths_E_N1_B_N2,
InBlockCopyClusterLengths_E_N1_B_N2,
InBlockCopyThreadClusterArrangeOrder,
InBlockCopySrcAccessOrder,
InBlockCopyDstAccessOrder,
InBlockCopySrcDataPerRead_B,
InBlockCopyDstDataPerWrite_N2,
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>{};
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
dim3(GridSize),
dim3(BlockSize),
0,
static_cast<T*>(in_nchw_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_kcyx_device_buf.GetDeviceBuffer()),
static_cast<T*>(out_nkhw_device_buf.GetDeviceBuffer()));
printf("Elapsed time : %f ms, %f TFlop/s\n",
time,
(float)calculate_convolution_flops(InDesc{}, WeiDesc{}, OutDesc{}) /
(std::size_t(1000) * 1000 * 1000) / time);
usleep(std::min(time * 1000, float(10000)));
}
out_nkhw_device_buf.FromDevice(out_nkhw.mData.data());
}

159
driver/include/device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw.hpp Normal file
View File

@@ -0,0 +1,159 @@
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "tensor.hpp"
#include "gridwise_convolution_kernel_wrapper.hpp"
#include "gridwise_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw_lds_double_buffer.hpp"
using namespace ck;
template <class T,
class InDesc,
class WeiDesc,
class OutDesc,
class ConvStrides,
class ConvDilations>
void device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw(InDesc,
const Tensor<T>& in_nchw,
WeiDesc,
const Tensor<T>& wei_kcyx,
OutDesc,
Tensor<T>& out_nkhw,
ConvStrides,
ConvDilations,
index_t nrepeat)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_nchw_desc = InDesc{};
constexpr auto wei_kcyx_desc = WeiDesc{};
constexpr auto out_nkhw_desc = OutDesc{};
constexpr index_t Hi = in_nchw_desc.GetLength(I2);
constexpr index_t Wi = in_nchw_desc.GetLength(I3);
constexpr index_t N = out_nkhw_desc.GetLength(I0);
constexpr index_t Ho = out_nkhw_desc.GetLength(I2);
constexpr index_t Wo = out_nkhw_desc.GetLength(I3);
constexpr index_t K = wei_kcyx_desc.GetLength(I0);
constexpr index_t C = wei_kcyx_desc.GetLength(I1);
constexpr index_t Y = wei_kcyx_desc.GetLength(I2);
constexpr index_t X = wei_kcyx_desc.GetLength(I3);
std::size_t data_sz = sizeof(T);
DeviceMem in_nchw_device_buf(data_sz * in_nchw.mDesc.GetElementSpace());
DeviceMem wei_kcyx_device_buf(data_sz * wei_kcyx.mDesc.GetElementSpace());
DeviceMem out_nkhw_device_buf(data_sz * out_nkhw.mDesc.GetElementSpace());
in_nchw_device_buf.ToDevice(in_nchw.mData.data());
wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
constexpr index_t N1 = 2;
constexpr index_t N2 = 4;
constexpr index_t B = (N * Ho * Wo) / (N1 * N2);
#if 1
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 16;
constexpr index_t KPerBlock = 128;
constexpr index_t EPerBlock = 8;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 4;
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopySubLengths_E_N1_B_N2 = Sequence<1, 1, 1, 4>;
using InBlockCopyClusterLengths_E_N1_B_N2 = Sequence<8, 2, 16, 1>;
using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopySrcAccessOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopyDstAccessOrder = Sequence<0, 1, 2, 3>; // [E, N1, B, N2]
constexpr index_t InBlockCopySrcDataPerRead_B = 1;
constexpr index_t InBlockCopyDstDataPerWrite_N2 = 4;
using WeiBlockCopySubLengths_E_K = Sequence<4, 1>;
using WeiBlockCopyClusterLengths_E_K = Sequence<2, 128>;
using WeiBlockCopyThreadClusterArrangeOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopySrcAccessOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopyDstAccessOrder = Sequence<0, 1>; // [E, K]
constexpr index_t WeiBlockCopySrcDataPerRead_E = 4;
constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
#endif
constexpr index_t GridSize =
((B + BPerBlock - 1) / BPerBlock) * ((K + KPerBlock - 1) / KPerBlock);
printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
for(index_t i = 0; i < nrepeat; ++i)
{
constexpr auto gridwise_conv =
GridwiseConvolutionImplicitGemm_v4r2_nchw_kcyx_nkhw_lds_double_buffer<
GridSize,
BlockSize,
T,
decltype(in_nchw_desc),
decltype(wei_kcyx_desc),
decltype(out_nkhw_desc),
ConvStrides,
ConvDilations,
BPerBlock,
KPerBlock,
EPerBlock,
N1,
N2,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB,
InBlockCopySubLengths_E_N1_B_N2,
InBlockCopyClusterLengths_E_N1_B_N2,
InBlockCopyThreadClusterArrangeOrder,
InBlockCopySrcAccessOrder,
InBlockCopyDstAccessOrder,
InBlockCopySrcDataPerRead_B,
InBlockCopyDstDataPerWrite_N2,
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>{};
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
dim3(GridSize),
dim3(BlockSize),
0,
static_cast<T*>(in_nchw_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_kcyx_device_buf.GetDeviceBuffer()),
static_cast<T*>(out_nkhw_device_buf.GetDeviceBuffer()));
printf("Elapsed time : %f ms, %f TFlop/s\n",
time,
(float)calculate_convolution_flops(InDesc{}, WeiDesc{}, OutDesc{}) /
(std::size_t(1000) * 1000 * 1000) / time);
usleep(std::min(time * 1000, float(10000)));
}
out_nkhw_device_buf.FromDevice(out_nkhw.mData.data());
}