ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 09:16:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

initial implementation for nchw v4r4 padding

Browse Source
This commit is contained in:
Chao Liu
2019-09-15 16:31:54 -05:00
parent 53094f7fae
commit 2c93b3057d
11 changed files with 721 additions and 503 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp
View File

@@ -100,10 +100,8 @@ struct GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw
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),
static_assert((Ho == 1 || ConvStrideW % InBlockCopySrcDataPerRead_B == 0) &&
(X == 1 || ConvDilationW % InBlockCopySrcDataPerRead_B == 0),
"wrong! aligment requirement for vectorized global load of input tensor will "
"be violated");

6
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer.hpp
View File

@@ -100,10 +100,8 @@ struct GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer
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),
static_assert((Ho == 1 || ConvStrideW % InBlockCopySrcDataPerRead_B == 0) &&
(X == 1 || ConvDilationW % InBlockCopySrcDataPerRead_B == 0),
"wrong! aligment requirement for vectorized global load of input tensor will "
"be violated");

6
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded.hpp
View File

@@ -107,10 +107,8 @@ struct GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_padded
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),
static_assert((Ho == 1 || ConvStrideW % InBlockCopySrcDataPerRead_B == 0) &&
(X == 1 || ConvDilationW % InBlockCopySrcDataPerRead_B == 0),
"wrong! aligment requirement for vectorized global load of input tensor will "
"be violated");

4
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp
View File

@@ -83,7 +83,9 @@ struct GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw
constexpr index_t E = C * Y * X;
constexpr index_t B = N * Ho * Wo;
static_assert((X == 1 || ConvDilationW % InBlockCopyDataPerAccess_B == 0),
// sanity-check for vectorized memory load
static_assert((Ho == 1 || ConvStrideW % InBlockCopyDataPerAccess_B == 0) &&
(X == 1 || ConvDilationW % InBlockCopyDataPerAccess_B == 0),
"wrong! aligment requirement for vectorized global load of input tensor will "
"be violated");

4
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_lds_double_buffer.hpp
View File

@@ -83,7 +83,9 @@ struct GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw_lds_double_buffer
constexpr index_t E = C * Y * X;
constexpr index_t B = N * Ho * Wo;
static_assert((X == 1 || ConvDilationW % InBlockCopyDataPerAccess_B == 0),
// sanity-check for vectorized memory load
static_assert((Ho == 1 || ConvStrideW % InBlockCopyDataPerAccess_B == 0) &&
(X == 1 || ConvDilationW % InBlockCopyDataPerAccess_B == 0),
"wrong! aligment requirement for vectorized global load of input tensor will "
"be violated");

457
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_padded.hpp Normal file
View File

@@ -0,0 +1,457 @@
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R4_NCHW_KCYX_NKHW_PADDED_HPP
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R4_NCHW_KCYX_NKHW_PADDED_HPP
#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 {
// B = merge(N, Ho, Wo)
template <index_t GridSize,
index_t BlockSize,
typename Float,
typename InGlobalDesc,
typename WeiGlobalDesc,
typename OutGlobalDesc,
typename ConvStrides,
typename ConvDilations,
typename LeftPads,
typename RightPads,
index_t BPerBlock,
index_t KPerBlock,
index_t EPerBlock,
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,
typename InBlockCopySubLengths_E_B,
typename InBlockCopyClusterLengths_E_B,
typename InBlockCopyThreadClusterArrangeOrder,
typename InBlockCopySrcAccessOrder,
typename InBlockCopyDstAccessOrder,
index_t InBlockCopyDataPerAccess_B,
typename WeiBlockCopySubLengths_E_K,
typename WeiBlockCopyClusterLengths_E_K,
typename WeiBlockCopyThreadClusterArrangeOrder,
typename WeiBlockCopySrcAccessOrder,
typename WeiBlockCopyDstAccessOrder,
index_t WeiBlockCopySrcDataPerRead_E,
index_t WeiBlockCopyDstDataPerWrite_K,
index_t OutThreadCopyDataPerAccess_B>
struct GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw_padded
{
#if 1
__device__ void Run(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global) const
{
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_hi_wi_global_desc =
make_native_tensor_descriptor(InGlobalDesc::GetLengths(), InGlobalDesc::GetStrides());
constexpr auto wei_k_c_y_x_global_desc =
make_native_tensor_descriptor(WeiGlobalDesc::GetLengths(), WeiGlobalDesc::GetStrides());
constexpr auto out_n_k_ho_wo_global_desc =
make_native_tensor_descriptor(OutGlobalDesc::GetLengths(), OutGlobalDesc::GetStrides());
constexpr index_t N = in_n_c_hi_wi_global_desc.GetLength(I0);
constexpr index_t C = in_n_c_hi_wi_global_desc.GetLength(I1);
constexpr index_t Hi = in_n_c_hi_wi_global_desc.GetLength(I2);
constexpr index_t Wi = in_n_c_hi_wi_global_desc.GetLength(I3);
constexpr index_t K = out_n_k_ho_wo_global_desc.GetLength(I1);
constexpr index_t Ho = out_n_k_ho_wo_global_desc.GetLength(I2);
constexpr index_t Wo = out_n_k_ho_wo_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];
constexpr index_t E = C * Y * X;
constexpr index_t B = N * Ho * Wo;
// sanity-check for vectorized memory load
static_assert((Ho == 1 || ConvStrideW % InBlockCopyDataPerAccess_B == 0) &&
(X == 1 || ConvDilationW % InBlockCopyDataPerAccess_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 % 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
// global mem
constexpr auto in_n_c_hip_wip_global_desc = transform_tensor_descriptor(
in_n_c_hi_wi_global_desc,
make_tuple(
PassThrough<N>{}, PassThrough<C>{}, Pad<Sequence<Hi, Wi>, LeftPads, RightPads>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}));
constexpr auto in_n_c_y_ho_x_wo_global_desc = transform_tensor_descriptor(
in_n_c_hip_wip_global_desc,
make_tuple(PassThrough<N>{},
PassThrough<C>{},
Embed<Sequence<Y, Ho>, Sequence<ConvDilationH, ConvStrideH, 0>>{},
Embed<Sequence<X, Wo>, Sequence<ConvDilationW, ConvStrideW, 0>>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}));
constexpr auto in_e_b_global_desc = transform_tensor_descriptor(
in_n_c_y_ho_x_wo_global_desc,
make_tuple(Merge<Sequence<C, Y, X>>{}, Merge<Sequence<N, Ho, Wo>>{}),
make_tuple(Sequence<1, 2, 4>{}, Sequence<0, 3, 5>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
// LDS mem
// be careful of LDS alignment
constexpr auto in_e_b_block_desc =
make_native_tensor_descriptor_packed(Sequence<EPerBlock, BPerBlock>{});
// input blockwise copy
auto blockwise_in_copy =
BlockwiseGenericTensorSliceCopy_v4<BlockSize,
decltype(in_e_b_global_desc),
decltype(in_e_b_block_desc),
decltype(in_e_b_block_desc.GetLengths()),
InBlockCopySubLengths_E_B,
InBlockCopyClusterLengths_E_B,
InBlockCopyThreadClusterArrangeOrder,
InBlockCopySrcAccessOrder,
InBlockCopyDstAccessOrder,
1,
1,
InBlockCopyDataPerAccess_B,
InBlockCopyDataPerAccess_B>(
{0, b_block_data_on_global}, {0, 0});
// weight tensor
// global mem
constexpr auto wei_e_k_global_desc =
transform_tensor_descriptor(wei_k_c_y_x_global_desc,
make_tuple(Merge<Sequence<C, Y, X>>{}, PassThrough<K>{}),
make_tuple(Sequence<1, 2, 3>{}, Sequence<0>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
// LDS
// be careful of LDS alignment
constexpr auto wei_e_k_block_desc = make_native_tensor_descriptor_aligned(
Sequence<EPerBlock, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDstDataPerWrite_K, GemmDataPerReadA)>{});
// weight blockwise copy
auto blockwise_wei_copy =
BlockwiseGenericTensorSliceCopy_v4<BlockSize,
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,
0,
1,
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, BPerBlock] is in LDS
// c_mtx[KPerBlock, BPerBlock] is distributed among threads, and saved in
// register
constexpr auto a_e_k_block_mtx_desc = make_ConstantMatrixDescriptor(wei_e_k_block_desc);
constexpr auto b_e_b_block_mtx_desc = make_ConstantMatrixDescriptor(in_e_b_block_desc);
// sanity check
static_assert(
KPerBlock % (GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster) == 0 &&
BPerBlock % (GemmNPerThreadSubC * GemmNLevel0Cluster * GemmNLevel1Cluster) == 0,
"wrong!");
constexpr index_t GemmMRepeat =
KPerBlock / (GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster);
constexpr index_t GemmNRepeat =
BPerBlock / (GemmNPerThreadSubC * GemmNLevel0Cluster * GemmNLevel1Cluster);
// c_thread_mtx definition: this is a mess
// TODO:: more elegent way of defining c_thread_mtx
constexpr auto c_k0k1_b0b1_thread_mtx_desc = make_ConstantMatrixDescriptor_packed(
Number<GemmMRepeat * GemmMPerThreadSubC>{}, Number<GemmNRepeat * GemmNPerThreadSubC>{});
const auto blockwise_gemm = BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2<
BlockSize,
decltype(a_e_k_block_mtx_desc),
decltype(b_e_b_block_mtx_desc),
decltype(c_k0k1_b0b1_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(InBlockCopyDataPerAccess_B,
WeiBlockCopyDstDataPerWrite_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr index_t in_block_space =
math::integer_least_multiple(in_e_b_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[in_block_space];
__shared__ Float p_wei_block[wei_block_space];
// register allocation for output
Float p_out_thread[c_k0k1_b0b1_thread_mtx_desc.GetElementSpace()];
// zero out threadwise output
threadwise_matrix_set_zero(c_k0k1_b0b1_thread_mtx_desc, p_out_thread);
for(index_t e_block_data_begin = 0; e_block_data_begin < E; e_block_data_begin += EPerBlock)
{
blockwise_in_copy.Run(p_in_global, p_in_block);
blockwise_wei_copy.Run(p_wei_global, p_wei_block);
__syncthreads();
blockwise_gemm.Run(p_wei_block, p_in_block, p_out_thread);
__syncthreads();
blockwise_in_copy.MoveSrcSliceWindow(make_multi_index(EPerBlock, 0), True);
blockwise_wei_copy.MoveSrcSliceWindow(make_multi_index(EPerBlock, 0), True);
}
// copy output: register to global memory
{
// 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;
// src descriptor
constexpr auto out_k0_k1_b0_b1_thread_desc = make_native_tensor_descriptor_packed(
Sequence<GemmMRepeat, GemmMPerThreadSubC, GemmNRepeat, GemmNPerThreadSubC>{});
// dst descriptor
constexpr index_t K1 = GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster;
constexpr index_t B1 = GemmNPerThreadSubC * GemmNLevel0Cluster * GemmNLevel1Cluster;
constexpr index_t K0 = K / K1;
constexpr index_t B0 = B / B1;
constexpr auto out_k_b_global_desc = transform_tensor_descriptor(
out_n_k_ho_wo_global_desc,
make_tuple(PassThrough<K>{}, Merge<Sequence<N, Ho, Wo>>{}),
make_tuple(Sequence<1>{}, Sequence<0, 2, 3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
constexpr auto out_k0_k1_b0_b1_global_desc = transform_tensor_descriptor(
out_k_b_global_desc,
make_tuple(Unmerge<Sequence<K0, K1>>{}, Unmerge<Sequence<B0, B1>>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
// output threadwise copy
auto threadwise_out_copy = ThreadwiseGenericTensorSliceCopy_v4r2<
decltype(out_k0_k1_b0_b1_thread_desc),
decltype(out_k0_k1_b0_b1_global_desc),
decltype(out_k0_k1_b0_b1_thread_desc.GetLengths()),
arithmetic_sequence_gen<0, 4, 1>::type,
3,
OutThreadCopyDataPerAccess_B,
OutThreadCopyDataPerAccess_B>({0, 0, 0, 0},
{k_thread_data_on_global / K1,
k_thread_data_on_global % K1,
b_thread_data_on_global / B1,
b_thread_data_on_global % B1});
threadwise_out_copy.Run(p_out_thread, p_out_global);
}
}
#else
__device__ void Run(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global) const
{
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_hi_wi_global_desc =
make_native_tensor_descriptor(InGlobalDesc::GetLengths(), InGlobalDesc::GetStrides());
constexpr auto wei_k_c_y_x_global_desc =
make_native_tensor_descriptor(WeiGlobalDesc::GetLengths(), WeiGlobalDesc::GetStrides());
constexpr auto out_n_k_ho_wo_global_desc =
make_native_tensor_descriptor(OutGlobalDesc::GetLengths(), OutGlobalDesc::GetStrides());
constexpr index_t N = in_n_c_hi_wi_global_desc.GetLength(I0);
constexpr index_t C = in_n_c_hi_wi_global_desc.GetLength(I1);
constexpr index_t Hi = in_n_c_hi_wi_global_desc.GetLength(I2);
constexpr index_t Wi = in_n_c_hi_wi_global_desc.GetLength(I3);
constexpr index_t K = out_n_k_ho_wo_global_desc.GetLength(I1);
constexpr index_t Ho = out_n_k_ho_wo_global_desc.GetLength(I2);
constexpr index_t Wo = out_n_k_ho_wo_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];
constexpr index_t E = C * Y * X;
constexpr index_t B = N * Ho * Wo;
// sanity-check for vectorized memory load
static_assert((Ho == 1 || ConvStrideW % InBlockCopyDataPerAccess_B == 0) &&
(X == 1 || ConvDilationW % InBlockCopyDataPerAccess_B == 0),
"wrong! aligment requirement for vectorized global load of input tensor will "
"be violated");
// input tensor
constexpr auto in_n_c_hip_wip_global_desc = transform_tensor_descriptor(
in_n_c_hi_wi_global_desc,
make_tuple(
PassThrough<N>{}, PassThrough<C>{}, Pad<Sequence<Hi, Wi>, LeftPads, RightPads>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}));
constexpr auto in_n_c_y_ho_x_wo_global_desc = transform_tensor_descriptor(
in_n_c_hip_wip_global_desc,
make_tuple(PassThrough<N>{},
PassThrough<C>{},
Embed<Sequence<Y, Ho>, Sequence<ConvDilationH, ConvStrideH, 0>>{},
Embed<Sequence<X, Wo>, Sequence<ConvDilationW, ConvStrideW, 0>>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}));
constexpr auto in_e_b_global_desc = transform_tensor_descriptor(
in_n_c_y_ho_x_wo_global_desc,
make_tuple(Merge<Sequence<C, Y, X>>{}, Merge<Sequence<N, Ho, Wo>>{}),
make_tuple(Sequence<1, 2, 4>{}, Sequence<0, 3, 5>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
// output tensor
constexpr auto out_k_b_global_desc =
transform_tensor_descriptor(out_n_k_ho_wo_global_desc,
make_tuple(PassThrough<K>{}, Merge<Sequence<N, Ho, Wo>>{}),
make_tuple(Sequence<1>{}, Sequence<0, 2, 3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
constexpr index_t K1 = GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster;
constexpr index_t B1 = GemmNPerThreadSubC * GemmNLevel0Cluster * GemmNLevel1Cluster;
constexpr index_t K0 = K / K1;
constexpr index_t B0 = B / B1;
constexpr auto out_k0_k1_b0_b1_global_desc = transform_tensor_descriptor(
out_k_b_global_desc,
make_tuple(Unmerge<Sequence<K0, K1>>{}, Unmerge<Sequence<B0, B1>>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
#if 1
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_tensor_descriptor("in_e_b_global_desc: ", in_e_b_global_desc);
print_tensor_descriptor("in_n_c_y_ho_x_wo_global_desc: ", in_n_c_y_ho_x_wo_global_desc);
print_tensor_descriptor("in_n_c_hip_wip_global_desc: ", in_n_c_hip_wip_global_desc);
print_tensor_descriptor("in_n_c_hi_wi_global_desc: ", in_n_c_hi_wi_global_desc);
auto coord3 = make_tensor_coordinate_v2(in_e_b_global_desc, {1, 1});
auto idx3 = coord3.GetIndex();
auto idx2 = coord3.GetLowerCoordinate().GetIndex();
auto idx1 = coord3.GetLowerCoordinate().GetLowerCoordinate().GetIndex();
auto idx0 =
coord3.GetLowerCoordinate().GetLowerCoordinate().GetLowerCoordinate().GetIndex();
print_array("idx3: ", idx3);
print_array("idx2: ", idx2);
print_array("idx1: ", idx1);
print_array("idx0: ", idx0);
}
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_tensor_descriptor("out_k0_k1_b0_b1_global_desc: ", out_k0_k1_b0_b1_global_desc);
print_tensor_descriptor("out_k_b_global_desc: ", out_k_b_global_desc);
print_tensor_descriptor("out_n_k_ho_wo_global_desc: ", out_n_k_ho_wo_global_desc);
auto coord2 = make_tensor_coordinate_v2(out_k0_k1_b0_b1_global_desc, {1, 1, 1, 1});
auto idx2 = coord2.GetIndex();
auto idx1 = coord2.GetLowerCoordinate().GetIndex();
auto idx0 = coord2.GetLowerCoordinate().GetLowerCoordinate().GetIndex();
print_array("idx2: ", idx2);
print_array("idx1: ", idx1);
print_array("idx0: ", idx0);
}
#endif
}
#endif
};
} // namespace ck
#endif

6
composable_kernel/include/tensor_description/tensor_coordinate_v2.hpp
View File

@@ -190,14 +190,16 @@ struct TensorCoordinate_v2
__host__ __device__ static constexpr auto
MakeDummyTensorCoordinate(NativeTensorDescriptor<Ts...>)
{
return NativeTensorCoordinate<NativeTensorDescriptor<Ts...>>();
return NativeTensorCoordinate<NativeTensorDescriptor<Ts...>>(
make_zero_array<index_t, TensorDesc::GetNumOfDimension()>());
}
template <typename... Ts>
__host__ __device__ static constexpr auto
MakeDummyTensorCoordinate(TransformedTensorDescriptor<Ts...>)
{
return TransformedTensorCoordinate<TransformedTensorDescriptor<Ts...>>();
return TransformedTensorCoordinate<TransformedTensorDescriptor<Ts...>>(
make_zero_array<index_t, TensorDesc::GetNumOfDimension()>());
}
public:

22
composable_kernel/include/tensor_description/tensor_descriptor.hpp
View File

@@ -187,8 +187,28 @@ struct TransformedTensorDescriptor
nTransform == UpDimensionIds::Size(),
"wrong! # of transformations not the same");
// TODO: sanity check: LowDimensionIds should include all low-dimensions,
// sanity check:
// LowDimensionIds should include all low-dimensions,
// UpDimensionIds should include all up-dimensions
using mingled_up_dimension_ids =
decltype(unpack(lambda_merge_sequences{}, UpDimensionIds{}));
using sorted_up_dimension_ids =
typename sequence_sort<mingled_up_dimension_ids, math::less<index_t>>::type;
static_assert(sorted_up_dimension_ids::Size() == nDimUp &&
is_valid_sequence_map<sorted_up_dimension_ids>{},
"wrong! UpDimensionIds is not configured correctly");
using mingled_low_dimension_ids =
decltype(unpack(lambda_merge_sequences{}, LowDimensionIds{}));
using sorted_low_dimension_ids =
typename sequence_sort<mingled_low_dimension_ids, math::less<index_t>>::type;
static_assert(sorted_low_dimension_ids::Size() == nDimLow &&
is_valid_sequence_map<sorted_low_dimension_ids>{},
"wrong! LowDimensionIds is not configured correctly");
// TODO: sanity check: while a up-dimension could be associated with multille
// transformation, a low-dimension should be associated with only one transformation

225
driver/include/device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_padded.hpp Normal file
View File

@@ -0,0 +1,225 @@
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "tensor.hpp"
#include "gridwise_convolution_kernel_wrapper.hpp"
#include "gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_padded.hpp"
template <class T,
class InDesc,
class WeiDesc,
class OutDesc,
class ConvStrides,
class ConvDilations,
class LeftPads,
class RightPads>
void device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_padded(InDesc,
const Tensor<T>& in_nchw,
WeiDesc,
const Tensor<T>& wei_kcyx,
OutDesc,
Tensor<T>& out_nkhw,
ConvStrides,
ConvDilations,
LeftPads,
RightPads,
index_t nrepeat)
{
using namespace ck;
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 N = out_nkhw_desc.GetLength(I0);
constexpr index_t K = out_nkhw_desc.GetLength(I1);
constexpr index_t Ho = out_nkhw_desc.GetLength(I2);
constexpr index_t Wo = out_nkhw_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());
#if 1
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 128;
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_B = Sequence<4, 1>;
using InBlockCopyClusterLengths_E_B = Sequence<2, 128>;
using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1>; // [E, B]
using InBlockCopySrcAccessOrder = Sequence<0, 1>; // [E, B]
using InBlockCopyDstAccessOrder = Sequence<0, 1>; // [E, B]
constexpr index_t InBlockCopyDataPerAccess_B = 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;
constexpr index_t OutThreadCopyDataPerAccess_B = 1;
#elif 1
// 1x1 filter, 8x8 image
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 128;
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_B = Sequence<1, 4>;
using InBlockCopyClusterLengths_E_B = Sequence<8, 32>;
using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1>; // [E, B]
using InBlockCopySrcAccessOrder = Sequence<0, 1>; // [E, B]
using InBlockCopyDstAccessOrder = Sequence<0, 1>; // [E, B]
constexpr index_t InBlockCopyDataPerAccess_B = 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;
constexpr index_t OutThreadCopyDataPerAccess_B = 4;
#elif 0
// 1x1 filter, 14x14 image
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 128;
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_B = Sequence<2, 2>;
using InBlockCopyClusterLengths_E_B = Sequence<4, 64>;
using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1>; // [E, B]
using InBlockCopySrcAccessOrder = Sequence<0, 1>; // [E, B]
using InBlockCopyDstAccessOrder = Sequence<0, 1>; // [E, B]
constexpr index_t InBlockCopyDataPerAccess_B = 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;
constexpr index_t OutThreadCopyDataPerAccess_B = 2;
#endif
constexpr index_t B = N * Ho * Wo;
constexpr index_t GridSize =
((B + BPerBlock - 1) / BPerBlock) * ((K + KPerBlock - 1) / KPerBlock);
printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
constexpr auto gridwise_conv = GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw_padded<
GridSize,
BlockSize,
T,
decltype(in_nchw_desc),
decltype(wei_kcyx_desc),
decltype(out_nkhw_desc),
ConvStrides,
ConvDilations,
LeftPads,
RightPads,
BPerBlock,
KPerBlock,
EPerBlock,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB,
InBlockCopySubLengths_E_B,
InBlockCopyClusterLengths_E_B,
InBlockCopyThreadClusterArrangeOrder,
InBlockCopySrcAccessOrder,
InBlockCopyDstAccessOrder,
InBlockCopyDataPerAccess_B,
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K,
OutThreadCopyDataPerAccess_B>{};
for(index_t i = 0; i < nrepeat; ++i)
{
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());
}

1
driver/src/driver.cpp
View File

@@ -19,6 +19,7 @@
//#include "device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw.hpp"
//#include "device_convolution_implicit_gemm_v4r3_nchw_kcyx_nkhw.hpp"
#include "device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp"
#include "device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_padded.hpp"
struct GeneratorTensor_1
{

486
driver/src/driver.cu
View File

@@ -1,486 +0,0 @@
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include "config.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "device.hpp"
#include "conv_common.hpp"
#include "host_conv.hpp"
#include "device_convolution_direct_v2_nchw_kcyx_nkhw.hpp"
#include "device_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hpp"
#include "device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded.hpp"
//#include "device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw.hpp"
//#include "device_convolution_implicit_gemm_v2_chwn_cyxk_khwn.hpp"
//#include "device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp"
#include "device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp"
#include "device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded.hpp"
//#include "device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw.hpp"
//#include "device_convolution_implicit_gemm_v4r3_nchw_kcyx_nkhw.hpp"
#include "device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp"
struct GeneratorTensor_1
{
template <class... Is>
double operator()(Is... is)
{
return 1;
}
};
struct GeneratorTensor_2
{
int min_value = 0;
int max_value = 1;
template <class... Is>
double operator()(Is...)
{
return (std::rand() % (max_value - min_value)) + min_value;
}
};
struct GeneratorTensor_3
{
template <class... Is>
double operator()(Is... is)
{
std::array<index_t, sizeof...(Is)> dims = {{static_cast<index_t>(is)...}};
auto f_acc = [](auto a, auto b) { return 100 * a + b; };
return std::accumulate(dims.begin(), dims.end(), index_t(0), f_acc);
}
};
struct GeneratorTensor_Checkboard
{
template <class... Ts>
double operator()(Ts... Xs) const
{
std::array<index_t, sizeof...(Ts)> dims = {{Xs...}};
return std::accumulate(dims.begin(),
dims.end(),
true,
[](bool init, index_t x) -> int { return init != (x % 2); })
? 1
: -1;
}
};
int main(int argc, char* argv[])
{
using namespace ck;
#if 0
constexpr index_t N = 32;
constexpr index_t C = 8;
constexpr index_t HI = 1;
constexpr index_t WI = 1;
constexpr index_t K = 128;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<1, 1>;
using RightPads = Sequence<0, 0>;
#elif 1
// 3x3, 34x34
constexpr index_t N = 64;
constexpr index_t C = 256;
constexpr index_t HI = 34;
constexpr index_t WI = 34;
constexpr index_t K = 128;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 1x1 filter, 8x8 image
// cudnn@V100 68%, ck@V100 72%, ck@P100 52%, ck@VII 42%
constexpr index_t N = 64;
constexpr index_t C = 1536;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 256;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 1x1 filter, 8x8 image
// cudnn@V100 77%, ck@V100 76%, ck@P100 79%, ck@VII 51%
constexpr index_t N = 128;
constexpr index_t C = 2048;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 384;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 1x1 filter, 7x7 image
// cudnn@V100 82%, ck@V100 76%, ck@P100 67%, ck@VII 64%
constexpr index_t N = 128;
constexpr index_t C = 832;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
constexpr index_t K = 384;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 1x1 filter, 8x8 image
// cudnn@V100 83%, ck@V100 75%, ck@P100 78%, ck@VII 65%
constexpr index_t N = 128;
constexpr index_t C = 1280;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 384;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 1x1 filter, 14x14 image
// cudnn@V100 62%, ck@V100 68%, ck@P100 70%, ck@VII 50%
constexpr index_t N = 128;
constexpr index_t C = 512;
constexpr index_t HI = 14;
constexpr index_t WI = 14;
constexpr index_t K = 128;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 1x1 filter, 8x8 image
// cudnn@V100 74%, ck@V100 57%, ck@P100 78%, ck@VII 61%
constexpr index_t N = 64;
constexpr index_t C = 1536;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 384;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 1x1 filter, 28x28 image
// cudnn@V100 86%, ck@V100 84%, ck@P100 80%, ck@VII 69%
constexpr index_t N = 128;
constexpr index_t C = 256;
constexpr index_t HI = 28;
constexpr index_t WI = 28;
constexpr index_t K = 128;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 1x1 filter, 7x7 image
// cudnn@V100 71%, ck@V100 55%, ck@P100 70%, ck@VII 62%
constexpr index_t N = 128;
constexpr index_t C = 832;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
constexpr index_t K = 256;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 3x3 filter, 2x2 stride, 35x35 input, 17x17 output
// cudnn@V100 90%, ck@V100 93%, ck@P100 83%, ck@VII 81%
constexpr index_t N = 128;
constexpr index_t C = 288;
constexpr index_t HI = 35;
constexpr index_t WI = 35;
constexpr index_t K = 384;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<2, 2>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 1
// 1x1 filter, 17x17 input
// cudnn@V100 81%, ck@V100 76%, ck@P100 70%, ck@VII 76%
constexpr index_t N = 128;
constexpr index_t C = 768;
constexpr index_t HI = 17;
constexpr index_t WI = 17;
constexpr index_t K = 128;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 1x1 filter, 14x14 image
// cudnn@V100 73%, ck@V100 71%, ck@P100 70%, ck@VII 64%
constexpr index_t N = 128;
constexpr index_t C = 528;
constexpr index_t HI = 14;
constexpr index_t WI = 14;
constexpr index_t K = 128;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 1x1 filter, 14x14 image
// cudnn@V100 73%, ck@V100 72%, ck@P100 79%, ck@VII 75%
constexpr index_t N = 128;
constexpr index_t C = 528;
constexpr index_t HI = 14;
constexpr index_t WI = 14;
constexpr index_t K = 256;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
// 1x1 filter, 7x7 image
// cudnn@V100 49%, ck@V100 50%, ck@P100 61%, ck@VII 52%
constexpr index_t N = 128;
constexpr index_t C = 832;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
constexpr index_t K = 128;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#endif
auto in_nchw_desc = make_ConstantTensorDescriptor_packed(Sequence<N, C, HI, WI>{});
auto wei_kcyx_desc = make_ConstantTensorDescriptor_packed(Sequence<K, C, Y, X>{});
auto out_nkhw_desc = get_convolution_with_padding_output_default_4d_tensor_descriptor(
in_nchw_desc, wei_kcyx_desc, ConvStrides{}, ConvDilations{}, LeftPads{}, RightPads{});
ostream_ConstantTensorDescriptor(in_nchw_desc, std::cout << "in_nchw_desc: ");
ostream_ConstantTensorDescriptor(wei_kcyx_desc, std::cout << "wei_kcyx_desc: ");
ostream_ConstantTensorDescriptor(out_nkhw_desc, std::cout << "out_nkhw_desc: ");
using in_data_t = float;
using out_data_t = float;
Tensor<in_data_t> in_nchw(make_TensorDescriptor(in_nchw_desc));
Tensor<in_data_t> wei_kcyx(make_TensorDescriptor(wei_kcyx_desc));
Tensor<out_data_t> out_nkhw_host(make_TensorDescriptor(out_nkhw_desc));
Tensor<out_data_t> out_nkhw_device(make_TensorDescriptor(out_nkhw_desc));
std::size_t num_thread = std::thread::hardware_concurrency();
if(argc != 3)
{
printf("arg1: do_verification, arg2: nrepeat\n");
exit(1);
}
bool do_verification = atoi(argv[1]);
index_t nrepeat = atoi(argv[2]);
if(do_verification)
{
#if 0
in_nchw.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
#elif 0
in_nchw.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
#elif 0
in_nchw.GenerateTensorValue(GeneratorTensor_3{}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
#elif 1
in_nchw.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
#elif 0
in_nchw.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
auto gen_wei = [](auto... is) {
return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
};
wei_kcyx.GenerateTensorValue(gen_wei, num_thread);
#endif
}
#if 0
device_convolution_direct_v2_nchw_kcyx_nkhw
(in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(
in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
LeftPads{},
RightPads{},
nrepeat);
#elif 0
device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(
in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v2_chwn_cyxk_khwn(
in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(
(in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
nrepeat);
#elif 0
device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
LeftPads{},
RightPads{},
nrepeat);
#elif 0
device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
nrepeat);
#elif 0
device_convolution_implicit_gemm_v4r3_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
nrepeat);
#elif 1
device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
nrepeat);
#elif 0
device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_padded(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
LeftPads{},
RightPads{},
nrepeat);
#endif
if(do_verification)
{
#if 1
if(Y == 3 && X == 3 && ConvStrides{}[0] == 1 && ConvStrides{}[1] == 1 &&
ConvDilations{}[0] == 1 && ConvDilations{}[1] == 1)
{
host_winograd_3x3_convolution(
in_nchw, wei_kcyx, out_nkhw_host, LeftPads{}, RightPads{});
}
else
#endif
{
host_direct_convolution(in_nchw,
wei_kcyx,
out_nkhw_host,
ConvStrides{},
ConvDilations{},
LeftPads{},
RightPads{});
}
check_error(out_nkhw_host, out_nkhw_device);
#if 0
LogRange(std::cout << "in_nchw : ", in_nchw.mData, ",") << std::endl;
LogRange(std::cout << "wei_kcyx: ", wei_kcyx.mData, ",") << std::endl;
LogRange(std::cout << "out_nkhw_host : ", out_nkhw_host.mData, ",") << std::endl;
LogRange(std::cout << "out_nkhw_device: ", out_nkhw_device.mData, ",") << std::endl;
#endif
}
}

1
driver/src/driver.cu Symbolic link
View File

@@ -0,0 +1 @@
driver.cpp