Skip to content

GitLab

Commit 1e3d69b9 authored by Chao Liu's avatar Chao Liu
Browse files

small test case for hip compiler

parent f0716f5b
Hide whitespace changes
Inline Side-by-side
Showing with 0 additions and 8180 deletions
composable_kernel/include/kernel_algorithm/gridwise_convolution_direct_v2_nchw_kcyx_nkhw.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_DIRECT_V2_NCHW_KCYX_NKHW
#define CK_GRIDWISE_CONVOLUTION_DIRECT_V2_NCHW_KCYX_NKHW
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "threadwise_tensor_slice_copy.hpp"
#include "threadwise_direct_convolution.hpp"
namespace ck {
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
index_t NPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
index_t NPerThread,
index_t KPerThread,
index_t CPerThread,
index_t HoPerThread,
index_t WoPerThread,
index_t InBlockCopyDataPerRead,
index_t WeiBlockCopyDataPerRead>
struct GridwiseConvolutionDirect_v2_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)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_nchw_global_desc = InGlobalDesc{};
constexpr auto wei_kcyx_global_desc = WeiGlobalDesc{};
constexpr auto out_nkhw_global_desc = OutGlobalDesc{};
constexpr index_t N = in_nchw_global_desc.GetLength(I0);
constexpr index_t K = wei_kcyx_global_desc.GetLength(I0);
constexpr index_t C = wei_kcyx_global_desc.GetLength(I1);
constexpr index_t Y = wei_kcyx_global_desc.GetLength(I2);
constexpr index_t X = wei_kcyx_global_desc.GetLength(I3);
constexpr auto wei_ke_global_desc = make_ConstantTensorDescriptor_packed(
Sequence<K, C * Y * X>{}); // 2d view of wei for blockwise copy
constexpr index_t HiPerBlock = HoPerBlock + Y - 1;
constexpr index_t WiPerBlock = WoPerBlock + X - 1;
constexpr auto in_nchw_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<NPerBlock, CPerBlock, HiPerBlock, WiPerBlock>{},
Number<InBlockCopyDataPerRead>{});
constexpr auto wei_ke_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<KPerBlock, CPerBlock * Y * X>{},
Number<WeiBlockCopyDataPerRead>{}); // 2d view of wei for blockwise copy
constexpr auto wei_kcyx_block_desc =
make_ConstantTensorDescriptor(Sequence<KPerBlock, CPerBlock, Y, X>{},
Sequence<wei_ke_block_desc.GetStride(I0), Y * X, X, 1>{});
// shared mem
constexpr index_t in_block_element_size =
in_nchw_block_desc.GetElementSpace(Number<InBlockCopyDataPerRead>{});
constexpr index_t wei_block_element_size =
wei_kcyx_block_desc.GetElementSpace(Number<WeiBlockCopyDataPerRead>{});
constexpr index_t max_align = InBlockCopyDataPerRead > WeiBlockCopyDataPerRead
? InBlockCopyDataPerRead
: WeiBlockCopyDataPerRead;
__shared__ Float
p_in_block[max_align * ((in_block_element_size + max_align - 1) / max_align)];
__shared__ Float
p_wei_block[max_align * ((wei_block_element_size + max_align - 1) / max_align)];
// threadwise tensors
constexpr index_t HiPerThread = HoPerThread + Y - 1;
constexpr index_t WiPerThread = WoPerThread + X - 1;
constexpr auto in_nchw_thread_block_desc = make_ConstantTensorDescriptor(
Sequence<NPerThread, CPerThread, HiPerThread, WiPerThread>{},
in_nchw_block_desc.GetStrides());
constexpr auto wei_kcyx_thread_block_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread, CPerThread, Y, X>{}, wei_kcyx_block_desc.GetStrides());
constexpr auto out_nkhw_thread_desc = get_convolution_output_default_4d_tensor_descriptor(
in_nchw_thread_block_desc, wei_kcyx_thread_block_desc);
// register
Float p_out_thread[out_nkhw_thread_desc.GetElementSpace()];
// divide block work
constexpr index_t NBlockWork =
(out_nkhw_global_desc.GetLength(I0) + NPerBlock - 1) / NPerBlock;
constexpr index_t KBlockWork =
(out_nkhw_global_desc.GetLength(I1) + KPerBlock - 1) / KPerBlock;
constexpr index_t HBlockWork =
(out_nkhw_global_desc.GetLength(I2) + HoPerBlock - 1) / HoPerBlock;
constexpr index_t WBlockWork =
(out_nkhw_global_desc.GetLength(I3) + WoPerBlock - 1) / WoPerBlock;
const index_t block_id = blockIdx.x;
index_t itmp = block_id;
const index_t n_block_work_id = itmp / (KBlockWork * HBlockWork * WBlockWork);
itmp -= n_block_work_id * (KBlockWork * HBlockWork * WBlockWork);
const index_t k_block_work_id = itmp / (HBlockWork * WBlockWork);
itmp -= k_block_work_id * (HBlockWork * WBlockWork);
const index_t h_block_work_id = itmp / WBlockWork;
const index_t w_block_work_id = itmp - h_block_work_id * WBlockWork;
const index_t n_block_data_begin = n_block_work_id * NPerBlock;
const index_t k_block_data_begin = k_block_work_id * KPerBlock;
const index_t ho_block_data_begin = h_block_work_id * HoPerBlock;
const index_t wo_block_data_begin = w_block_work_id * WoPerBlock;
const index_t hi_block_data_begin = ho_block_data_begin; // minus padding
const index_t wi_block_data_begin = wo_block_data_begin; // minus padding
// divide thread work
constexpr index_t NThreadWork = (NPerBlock + NPerThread - 1) / NPerThread;
constexpr index_t KThreadWork = (KPerBlock + KPerThread - 1) / KPerThread;
constexpr index_t HThreadWork = (HoPerBlock + HoPerThread - 1) / HoPerThread;
constexpr index_t WThreadWork = (WoPerBlock + WoPerThread - 1) / WoPerThread;
const index_t thread_id = get_thread_local_1d_id();
itmp = thread_id;
const index_t n_thread_work_id = itmp / (KThreadWork * HThreadWork * WThreadWork);
itmp -= n_thread_work_id * (KThreadWork * HThreadWork * WThreadWork);
const index_t k_thread_work_id = itmp / (HThreadWork * WThreadWork);
itmp -= k_thread_work_id * (HThreadWork * WThreadWork);
const index_t h_thread_work_id = itmp / WThreadWork;
const index_t w_thread_work_id = itmp - h_thread_work_id * WThreadWork;
const index_t n_thread_data_begin = n_thread_work_id * NPerThread;
const index_t k_thread_data_begin = k_thread_work_id * KPerThread;
const index_t ho_thread_data_begin = h_thread_work_id * HoPerThread;
const index_t wo_thread_data_begin = w_thread_work_id * WoPerThread;
const index_t hi_thread_data_begin = ho_thread_data_begin;
const index_t wi_thread_data_begin = wo_thread_data_begin;
constexpr auto blockwise_in_copy =
Blockwise4dTensorCopy1<BlockSize,
Float,
decltype(in_nchw_global_desc),
decltype(in_nchw_block_desc),
decltype(in_nchw_block_desc.GetLengths()),
InBlockCopyDataPerRead>{};
#if 0
constexpr auto blockwise_wei_copy =
Blockwise4dTensorCopy1<BlockSize,
Float,
decltype(wei_kcyx_global_desc),
decltype(wei_kcyx_block_desc),
decltype(wei_kcyx_block_desc.GetLengths()),
1>{};
#elif 1
const auto blockwise_wei_copy =
Blockwise2dTensorCopy3<BlockSize,
Float,
decltype(wei_ke_global_desc),
decltype(wei_ke_block_desc),
decltype(wei_ke_block_desc.GetLengths()),
WeiBlockCopyDataPerRead>({0, 0}, {0, 0});
#endif
// set threadwise output tensor to 0
threadwise_4d_tensor_set_zero(out_nkhw_thread_desc, p_out_thread);
for(index_t c_block_data_begin = 0; c_block_data_begin < C;
c_block_data_begin += CPerBlock, __syncthreads())
{
// copy input tensor to LDS
blockwise_in_copy.Run(
p_in_global +
in_nchw_global_desc.GetOffsetFromMultiIndex(n_block_data_begin,
c_block_data_begin,
hi_block_data_begin,
wi_block_data_begin),
p_in_block);
// copy weight tensor to LDS
blockwise_wei_copy.Run(p_wei_global +
wei_kcyx_global_desc.GetOffsetFromMultiIndex(
k_block_data_begin, c_block_data_begin, 0, 0),
p_wei_block);
__syncthreads();
for(index_t c_thread_data = 0; c_thread_data < CPerBlock; c_thread_data += CPerThread)
{
// threadwise convolution
#if 1
threadwise_direct_convolution_2(
in_nchw_thread_block_desc,
p_in_block +
in_nchw_block_desc.GetOffsetFromMultiIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
wei_kcyx_thread_block_desc,
p_wei_block +
wei_kcyx_block_desc.GetOffsetFromMultiIndex(
k_thread_data_begin, c_thread_data, 0, 0),
out_nkhw_thread_desc,
p_out_thread);
#elif 0
threadwise_direct_convolution_3(
in_nchw_thread_block_desc,
p_in_block +
in_nchw_block_desc.GetOffsetFromMultiIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
wei_kcyx_thread_block_desc,
p_wei_block +
wei_kcyx_block_desc.GetOffsetFromMultiIndex(
k_thread_data_begin, c_thread_data, 0, 0),
out_nkhw_thread_desc,
p_out_thread);
#endif
}
}
// copy output tensor from register to global mem
threadwise_tensor_slice_copy(out_nkhw_thread_desc,
p_out_thread,
out_nkhw_global_desc,
p_out_global +
out_nkhw_global_desc.GetOffsetFromMultiIndex(
n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
out_nkhw_thread_desc.GetLengths(),
Number<1>{});
}
};
} // namespace ck
#endif
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r1_chwn_cyxk_khwn.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R1_CHWN_CYXK_KHWN
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R1_CHWN_CYXK_KHWN
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "threadwise_tensor_slice_copy.hpp"
#include "threadwise_4d_tensor_op.hpp"
#include "blockwise_batched_gemm.hpp"
namespace ck {
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
index_t NPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
index_t NPerThread,
index_t KPerThread,
index_t HoPerThread,
index_t WoPerThread,
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 InBlockCopyClusterLengths_CHWN,
index_t InBlockCopyDataPerRead_N,
index_t WeiBlockCopyDataPerRead_K,
index_t OutThreadCopyDataPerWrite_N>
struct GridwiseConvolutionImplicitGemm_v1r1_chwn_cyxk_khwn
{
__device__ void Run(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global) const
{
// be careful of this assertion
static_assert(
NPerBlock % NPerThread == 0 &&
((GemmNPerThreadSubC <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0) ||
(GemmNPerThreadSubC >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0)),
"wrong!");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_c_h_w_n_global_desc = InGlobalDesc{};
constexpr auto wei_c_y_x_k_global_desc = WeiGlobalDesc{};
constexpr auto out_k_h_w_n_global_desc = OutGlobalDesc{};
constexpr index_t C = in_c_h_w_n_global_desc.GetLength(I0);
constexpr index_t K = out_k_h_w_n_global_desc.GetLength(I0);
constexpr index_t Ho = out_k_h_w_n_global_desc.GetLength(I1);
constexpr index_t Wo = out_k_h_w_n_global_desc.GetLength(I2);
constexpr index_t N = out_k_h_w_n_global_desc.GetLength(I3);
constexpr index_t Y = wei_c_y_x_k_global_desc.GetLength(I1);
constexpr index_t X = wei_c_y_x_k_global_desc.GetLength(I2);
constexpr index_t HiPerBlock = HoPerBlock + Y - 1;
constexpr index_t WiPerBlock = WoPerBlock + X - 1;
// divide block work: [K, Ho, Wo, N]
static_assert(N % NPerBlock == 0 && K % KPerBlock == 0 && C % CPerBlock == 0 &&
Ho % HoPerBlock == 0 && Wo % WoPerBlock == 0,
"wrong! cannot evenly divide work for workgroup ");
constexpr index_t KBlockWork = (K + KPerBlock - 1) / KPerBlock;
constexpr index_t HBlockWork = (Ho + HoPerBlock - 1) / HoPerBlock;
constexpr index_t WBlockWork = (Wo + WoPerBlock - 1) / WoPerBlock;
constexpr index_t NBlockWork = (N + NPerBlock - 1) / NPerBlock;
const index_t k_block_work_id = get_block_1d_id() / (HBlockWork * WBlockWork * NBlockWork);
index_t itmp = get_block_1d_id() - k_block_work_id * (HBlockWork * WBlockWork * NBlockWork);
const index_t h_block_work_id = itmp / (WBlockWork * NBlockWork);
itmp -= h_block_work_id * (WBlockWork * NBlockWork);
const index_t w_block_work_id = itmp / NBlockWork;
const index_t n_block_work_id = itmp - w_block_work_id * NBlockWork;
const index_t k_block_data_begin = k_block_work_id * KPerBlock;
const index_t ho_block_data_begin = h_block_work_id * HoPerBlock;
const index_t wo_block_data_begin = w_block_work_id * WoPerBlock;
const index_t n_block_data_begin = n_block_work_id * NPerBlock;
const index_t hi_block_data_begin = ho_block_data_begin;
const index_t wi_block_data_begin = wo_block_data_begin;
// flattend (2d) tensor view of gridwise weight
constexpr auto wei_cyx_k_global_desc =
make_ConstantTensorDescriptor(Sequence<C * Y * X, K>{});
// tensor view of blockwise input and weight in LDS
// be careful of alignment
constexpr index_t max_align = math::lcm(InBlockCopyDataPerRead_N,
WeiBlockCopyDataPerRead_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, HiPerBlock, WiPerBlock, NPerBlock>{},
Number<InBlockCopyDataPerRead_N>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with alignment
static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not meet");
constexpr auto wei_cyx_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock * Y * X, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
constexpr auto wei_c_y_x_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, Y, X, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
// tensor view of threadwise output in register
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
// blockwise copy
// input: format is [C, Hi, Wi, N]
const auto blockwise_in_copy =
#if 0
Blockwise4dTensorCopy1<BlockSize,
Float,
decltype(in_c_h_w_n_global_desc),
decltype(in_c_h_w_n_block_desc),
decltype(in_c_h_w_n_block_desc.GetLengths()),
InBlockCopyDataPerRead_N>{};
#else
Blockwise4dTensorCopy3<BlockSize,
Float,
decltype(in_c_h_w_n_global_desc),
decltype(in_c_h_w_n_block_desc),
decltype(in_c_h_w_n_block_desc.GetLengths()),
InBlockCopyClusterLengths_CHWN,
InBlockCopyDataPerRead_N>{};
#endif
// blockwise wei copy
// format is [CPerBlock*Y*X,KPerBlock]
const auto blockwise_wei_copy =
Blockwise2dTensorCopy3<BlockSize,
Float,
decltype(wei_cyx_k_global_desc),
decltype(wei_cyx_k_block_desc),
decltype(wei_cyx_k_block_desc.GetLengths()),
WeiBlockCopyDataPerRead_K>{};
// a series of blockwise batched GEMM
// C_matrix += transpose(A_matrix) * B_matrix
// A_matrix and B_matrix saved in LDS, C_matrix saved in register
// A_matrix[C,K] is a sub-matrix of wei_block[C,Y,X,K]
// B_matrix[C,Wo*N] is a sub-matrix of in_block[C,Hi,Wi,N]
// C_matrix[K,Wo*N] is a sub-matrix of out_block[K,Ho,Wo,N]
constexpr auto a_c_k_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
Number<KPerBlock>{},
Number<wei_c_y_x_k_block_desc.GetStride(I0)>{});
constexpr auto b_c_wn_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
Number<WoPerBlock * NPerBlock>{},
Number<in_c_h_w_n_block_desc.GetStride(I0)>{});
constexpr auto c_k_wn_thread_mtx_desc =
make_ConstantMatrixDescriptor(Number<KPerThread>{},
Number<WoPerThread * NPerThread>{},
Number<out_k_h_w_n_thread_desc.GetStride(I0)>{});
const auto blockwise_batch_gemm =
BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2<
BlockSize,
decltype(a_c_k_block_mtx_desc),
decltype(b_c_wn_block_mtx_desc),
decltype(c_k_wn_thread_mtx_desc),
0,
in_c_h_w_n_block_desc.GetStride(I1),
out_k_h_w_n_thread_desc.GetStride(I1),
HoPerBlock,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
HoPerThread,
GemmDataPerReadA,
GemmDataPerReadB>{};
// LDS: be careful of alignment
constexpr index_t in_block_space =
in_c_h_w_n_block_desc.GetElementSpace(Number<max_align>{});
constexpr index_t wei_block_space =
wei_c_y_x_k_block_desc.GetElementSpace(Number<max_align>{});
__shared__ Float p_in_block[in_block_space];
__shared__ Float p_wei_block[wei_block_space];
// register
// C++ lambda doesn't capture array, use pointer instead
Float p_out_thread_data[out_k_h_w_n_thread_desc.GetElementSpace()];
Float* const p_out_thread = p_out_thread_data;
// set threadwise output tensor to 0
threadwise_4d_tensor_set_zero(out_k_h_w_n_thread_desc, p_out_thread);
const Float* p_in_global_block_offset =
p_in_global +
in_c_h_w_n_global_desc.GetOffsetFromMultiIndex(
0, hi_block_data_begin, wi_block_data_begin, n_block_data_begin);
const Float* p_wei_global_block_offset =
p_wei_global +
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, 0, 0, k_block_data_begin);
for(index_t c_block_data_begin = 0; c_block_data_begin < C; c_block_data_begin += CPerBlock,
p_in_global_block_offset += CPerBlock * in_c_h_w_n_global_desc.GetStride(I0),
p_wei_global_block_offset += CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0),
__syncthreads())
{
#if 1
blockwise_in_copy.Run(p_in_global_block_offset, p_in_block);
blockwise_wei_copy.Run(p_wei_global_block_offset, p_wei_block);
#else
Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global_block_offset,
p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
p_wei_register_clipboard);
blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard, p_in_block);
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard, p_wei_block);
#endif
__syncthreads();
#pragma unroll
for(index_t y = 0; y < Y; ++y)
{
#pragma unroll
for(index_t x = 0; x < X; ++x)
{
#if 1
blockwise_batch_gemm.Run
#else
blockwise_batch_gemm.Run_amd_asm
#endif
(p_wei_block + wei_c_y_x_k_block_desc.GetOffsetFromMultiIndex(0, y, x, 0),
p_in_block + in_c_h_w_n_block_desc.GetOffsetFromMultiIndex(0, y, x, 0),
p_out_thread);
}
}
}
// output: register to global mem,
const auto c_thread_mtx_begin =
blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t k_thread_data_begin = c_thread_mtx_begin.row;
const index_t ho_thread_data_begin = c_thread_mtx_begin.batch;
const index_t wo_thread_data_begin = c_thread_mtx_begin.col / NPerBlock;
const index_t n_thread_data_begin = c_thread_mtx_begin.col % NPerBlock;
static_if<GemmNPerThreadSubC <= NPerBlock>{}([&](auto f_dummy) { // f_dummy do nothing but
// perfect forwarding.
// Using this trick to
// make this lambda a generic lambda, so it won't be compiled until
// instantiated
static_assert(
(f_dummy(GemmNPerThreadSubC) <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0),
"wrong!");
// output is a 10d tensor
constexpr index_t N2 = GemmNPerThreadSubC;
constexpr index_t N1 = NPerBlock / N2;
constexpr index_t W2 =
(GemmNLevel0Cluster * GemmNLevel1Cluster) / f_dummy(NPerBlock / GemmNPerThreadSubC);
constexpr index_t W1 = WoPerBlock / W2;
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc =
make_ConstantTensorDescriptor(Sequence<K / (K1 * K2),
K1,
K2,
Ho,
Wo / (W1 * W2),
W1,
W2,
N / f_dummy(N1 * N2),
N1,
N2>{});
constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread / K2, 1, K2, HoPerThread, 1, W1, 1, 1, 1, N2>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "out_10d_thread_desc");
print_ConstantTensorDescriptor(out_k_h_w_n_global_desc,
"out_k_h_w_n_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "out_10d_global_desc");
}
#endif
threadwise_tensor_slice_copy(out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_k_h_w_n_global_desc.GetOffsetFromMultiIndex(
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin,
n_block_data_begin + n_thread_data_begin),
out_10d_thread_desc.GetLengths(),
Number<OutThreadCopyDataPerWrite_N>{});
}).Else([&](auto f_dummy) {
static_assert(f_dummy(GemmNPerThreadSubC) >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0,
"wrong!");
// output is a 10d tensor
constexpr index_t N1 = NPerBlock;
constexpr index_t W3 = GemmNPerThreadSubC / NPerBlock;
constexpr index_t W2 = GemmNLevel0Cluster * GemmNLevel1Cluster;
constexpr index_t W1 = WoPerBlock / f_dummy(W2 * W3);
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc = make_ConstantTensorDescriptor(
Sequence<K / (K1 * K2), K1, K2, Ho, Wo / (W1 * W2 * W3), W1, W2, W3, N / N1, N1>{});
constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread / K2, 1, K2, HoPerThread, 1, W1, 1, W3, 1, N1>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "out_10d_thread_desc");
print_ConstantTensorDescriptor(out_k_h_w_n_global_desc,
"out_k_h_w_n_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "out_10d_global_desc");
for(index_t i = 0; i < 64; ++i)
{
printf("out %f, ", p_out_thread[i]);
}
}
#endif
threadwise_tensor_slice_copy(out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_k_h_w_n_global_desc.GetOffsetFromMultiIndex(
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin,
n_block_data_begin + n_thread_data_begin),
out_10d_thread_desc.GetLengths(),
Number<OutThreadCopyDataPerWrite_N>{});
});
}
};
} // namespace ck
#endif
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r2_chwn_cyxk_khwn.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R2_CHWN_CYXK_KHWN
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R2_CHWN_CYXK_KHWN
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_3d_tensor_op.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "threadwise_tensor_slice_copy.hpp"
#include "threadwise_4d_tensor_op.hpp"
#include "blockwise_batched_gemm.hpp"
namespace ck {
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
index_t NPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
index_t NPerThread,
index_t KPerThread,
index_t HoPerThread,
index_t WoPerThread,
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 InBlockCopyClusterLengths_CHWN,
index_t InBlockCopyDataPerRead_N,
index_t WeiBlockCopyDataPerRead_K,
index_t OutThreadCopyDataPerWrite_N>
struct GridwiseConvolutionImplicitGemm_v1r2_chwn_cyxk_khwn
{
__device__ void Run(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global) const
{
// be careful of this assertion
static_assert(
NPerBlock % NPerThread == 0 &&
((GemmNPerThreadSubC <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0) ||
(GemmNPerThreadSubC >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0)),
"wrong!");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_c_h_w_n_global_desc = InGlobalDesc{};
constexpr auto wei_c_y_x_k_global_desc = WeiGlobalDesc{};
constexpr auto out_k_h_w_n_global_desc = OutGlobalDesc{};
constexpr index_t C = in_c_h_w_n_global_desc.GetLength(I0);
constexpr index_t K = out_k_h_w_n_global_desc.GetLength(I0);
constexpr index_t Ho = out_k_h_w_n_global_desc.GetLength(I1);
constexpr index_t Wo = out_k_h_w_n_global_desc.GetLength(I2);
constexpr index_t N = out_k_h_w_n_global_desc.GetLength(I3);
constexpr index_t Y = wei_c_y_x_k_global_desc.GetLength(I1);
constexpr index_t X = wei_c_y_x_k_global_desc.GetLength(I2);
constexpr index_t HiPerBlock = HoPerBlock + Y - 1;
constexpr index_t WiPerBlock = WoPerBlock + X - 1;
// divide block work: [K, Ho, Wo, N]
static_assert(N % NPerBlock == 0 && K % KPerBlock == 0 && C % CPerBlock == 0 &&
Ho % HoPerBlock == 0 && Wo % WoPerBlock == 0,
"wrong! cannot evenly divide work for workgroup ");
constexpr index_t KBlockWork = (K + KPerBlock - 1) / KPerBlock;
constexpr index_t HBlockWork = (Ho + HoPerBlock - 1) / HoPerBlock;
constexpr index_t WBlockWork = (Wo + WoPerBlock - 1) / WoPerBlock;
constexpr index_t NBlockWork = (N + NPerBlock - 1) / NPerBlock;
const index_t k_block_work_id = get_block_1d_id() / (HBlockWork * WBlockWork * NBlockWork);
index_t itmp = get_block_1d_id() - k_block_work_id * (HBlockWork * WBlockWork * NBlockWork);
const index_t h_block_work_id = itmp / (WBlockWork * NBlockWork);
itmp -= h_block_work_id * (WBlockWork * NBlockWork);
const index_t w_block_work_id = itmp / NBlockWork;
const index_t n_block_work_id = itmp - w_block_work_id * NBlockWork;
const index_t k_block_data_begin = k_block_work_id * KPerBlock;
const index_t ho_block_data_begin = h_block_work_id * HoPerBlock;
const index_t wo_block_data_begin = w_block_work_id * WoPerBlock;
const index_t n_block_data_begin = n_block_work_id * NPerBlock;
const index_t hi_block_data_begin = ho_block_data_begin;
const index_t wi_block_data_begin = wo_block_data_begin;
// global tensor view
constexpr auto wei_c_x_k_global_desc =
make_ConstantTensorDescriptor(Sequence<C, X, K>{}, Sequence<Y * X * K, K, 1>{});
// LDS tensor view
// be careful of alignment
constexpr index_t max_align = math::lcm(InBlockCopyDataPerRead_N,
WeiBlockCopyDataPerRead_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, HoPerBlock, WiPerBlock, NPerBlock>{},
Number<InBlockCopyDataPerRead_N>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with alignment
static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not meet");
constexpr auto wei_c_x_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, X, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
// tensor view of threadwise output in register
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
// blockwise copy
// input: format is [C, Hi, Wi, N]
#if 1
const auto blockwise_in_copy =
Blockwise4dTensorCopy1<BlockSize,
Float,
decltype(in_c_h_w_n_global_desc),
decltype(in_c_h_w_n_block_desc),
decltype(in_c_h_w_n_block_desc.GetLengths()),
InBlockCopyDataPerRead_N>{};
#else
const auto blockwise_in_copy =
Blockwise4dTensorCopy3<BlockSize,
Float,
decltype(in_c_h_w_n_global_desc),
decltype(in_c_h_w_n_block_desc),
decltype(in_c_h_w_n_block_desc.GetLengths()),
InBlockCopyClusterLengths_CHWN,
InBlockCopyDataPerRead_N>{};
#endif
// blockwise wei copy
// format is [CPerBlock, X * KPerBlock]
const auto blockwise_wei_copy =
Blockwise3dTensorCopy1<BlockSize,
Float,
decltype(wei_c_x_k_global_desc),
decltype(wei_c_x_k_block_desc),
decltype(wei_c_x_k_block_desc.GetLengths()),
WeiBlockCopyDataPerRead_K>{};
// a series of blockwise batched GEMM
// C_matrix += transpose(A_matrix) * B_matrix
// A_matrix and B_matrix saved in LDS, C_matrix saved in register
// A_matrix[C,K] is a sub-matrix of wei_block[C,K]
// B_matrix[C,Wo*N] is a sub-matrix of in_block[C,Hi,Wi,N]
// C_matrix[K,Wo*N] is a sub-matrix of out_block[K,Ho,Wo,N]
constexpr auto a_c_k_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_c_x_k_block_desc.GetStride(I0)>{});
constexpr auto b_c_wn_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
Number<WoPerBlock * NPerBlock>{},
Number<in_c_h_w_n_block_desc.GetStride(I0)>{});
constexpr auto c_k_wn_thread_mtx_desc =
make_ConstantMatrixDescriptor(Number<KPerThread>{},
Number<WoPerThread * NPerThread>{},
Number<out_k_h_w_n_thread_desc.GetStride(I0)>{});
const auto blockwise_batch_gemm =
BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2<
BlockSize,
decltype(a_c_k_block_mtx_desc),
decltype(b_c_wn_block_mtx_desc),
decltype(c_k_wn_thread_mtx_desc),
0,
in_c_h_w_n_block_desc.GetStride(I1),
out_k_h_w_n_thread_desc.GetStride(I1),
HoPerBlock,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
HoPerThread,
GemmDataPerReadA,
GemmDataPerReadB>{};
// LDS: be careful of alignment
constexpr index_t in_block_space =
in_c_h_w_n_block_desc.GetElementSpace(Number<max_align>{});
constexpr index_t wei_block_space =
wei_c_x_k_block_desc.GetElementSpace(Number<max_align>{});
__shared__ Float p_in_block[in_block_space];
__shared__ Float p_wei_block[wei_block_space];
// register
// C++ lambda doesn't capture array, use pointer instead
Float p_out_thread_data[out_k_h_w_n_thread_desc.GetElementSpace()];
Float* const p_out_thread = p_out_thread_data;
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_c_h_w_n_global_desc, "in_c_h_w_n_global_desc");
print_ConstantTensorDescriptor(wei_c_y_x_k_global_desc, "wei_c_y_x_k_global_desc");
print_ConstantTensorDescriptor(in_c_h_w_n_block_desc, "in_c_h_w_n_block_desc");
print_ConstantTensorDescriptor(wei_c_x_k_block_desc, "wei_c_x_k_block_desc");
printf("in_block_space %u, wei_block_space %u\n", in_block_space, wei_block_space);
}
#endif
// set threadwise output tensor to 0
threadwise_4d_tensor_set_zero(out_k_h_w_n_thread_desc, p_out_thread);
#if 1
const Float* p_in_global_block_offset =
p_in_global +
in_c_h_w_n_global_desc.GetOffsetFromMultiIndex(
0, hi_block_data_begin, wi_block_data_begin, n_block_data_begin);
const Float* p_wei_global_block_offset =
p_wei_global +
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, 0, 0, k_block_data_begin);
for(index_t c_block_data_begin = 0; c_block_data_begin < C; c_block_data_begin += CPerBlock,
p_in_global_block_offset += CPerBlock * in_c_h_w_n_global_desc.GetStride(I0),
p_wei_global_block_offset += CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0))
{
for(index_t y = 0; y < Y; ++y)
{
blockwise_in_copy.Run(
p_in_global_block_offset +
in_c_h_w_n_global_desc.GetOffsetFromMultiIndex(0, y, 0, 0),
p_in_block);
blockwise_wei_copy.Run(
p_wei_global_block_offset +
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, 0, 0),
p_wei_block);
__syncthreads();
for(index_t x = 0; x < X; ++x)
{
blockwise_batch_gemm.Run(
p_wei_block + wei_c_x_k_block_desc.GetOffsetFromMultiIndex(0, x, 0),
p_in_block + in_c_h_w_n_block_desc.GetOffsetFromMultiIndex(0, 0, x, 0),
p_out_thread);
}
__syncthreads();
}
}
#else
// this use much more register, haven't figure out why?
for(index_t y = 0; y < Y; ++y)
{
const Float* p_in_global_block_offset =
p_in_global +
in_c_h_w_n_global_desc.GetOffsetFromMultiIndex(
0, hi_block_data_begin + y, wi_block_data_begin, n_block_data_begin);
const Float* p_wei_global_block_offset =
p_wei_global +
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, 0, k_block_data_begin);
for(index_t
c_block_data_begin = 0;
c_block_data_begin < C;
c_block_data_begin += CPerBlock,
p_in_global_block_offset += CPerBlock * in_c_h_w_n_global_desc.GetStride(I0),
p_wei_global_block_offset += CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0))
{
blockwise_in_copy.Run(p_in_global_block_offset, p_in_block);
blockwise_wei_copy.Run(p_wei_global_block_offset, p_wei_block);
__syncthreads();
for(index_t x = 0; x < X; ++x)
{
blockwise_batch_gemm.Run(
p_wei_block + wei_c_x_k_block_desc.GetOffsetFromMultiIndex(0, x, 0),
p_in_block + in_c_h_w_n_block_desc.GetOffsetFromMultiIndex(0, 0, x, 0),
p_out_thread);
}
__syncthreads();
}
}
#endif
// output: register to global mem,
const auto c_thread_mtx_begin =
blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t k_thread_data_begin = c_thread_mtx_begin.row;
const index_t ho_thread_data_begin = c_thread_mtx_begin.batch;
const index_t wo_thread_data_begin = c_thread_mtx_begin.col / NPerBlock;
const index_t n_thread_data_begin = c_thread_mtx_begin.col % NPerBlock;
static_if<GemmNPerThreadSubC <= NPerBlock>{}([&](auto f_dummy) { // f_dummy do nothing but
// perfect forwarding.
// Using this trick to
// make this lambda a generic lambda, so it won't be compiled until
// instantiated
static_assert(
(f_dummy(GemmNPerThreadSubC) <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0),
"wrong!");
// output is a 10d tensor
constexpr index_t N2 = GemmNPerThreadSubC;
constexpr index_t N1 = NPerBlock / N2;
constexpr index_t W2 =
(GemmNLevel0Cluster * GemmNLevel1Cluster) / f_dummy(NPerBlock / GemmNPerThreadSubC);
constexpr index_t W1 = WoPerBlock / W2;
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc =
make_ConstantTensorDescriptor(Sequence<K / (K1 * K2),
K1,
K2,
Ho,
Wo / (W1 * W2),
W1,
W2,
N / f_dummy(N1 * N2),
N1,
N2>{});
constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread / K2, 1, K2, HoPerThread, 1, W1, 1, 1, 1, N2>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "out_10d_thread_desc");
print_ConstantTensorDescriptor(out_k_h_w_n_global_desc,
"out_k_h_w_n_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "out_10d_global_desc");
}
#endif
threadwise_tensor_slice_copy(out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_k_h_w_n_global_desc.GetOffsetFromMultiIndex(
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin,
n_block_data_begin + n_thread_data_begin),
out_10d_thread_desc.GetLengths(),
Number<OutThreadCopyDataPerWrite_N>{});
}).Else([&](auto f_dummy) {
static_assert(f_dummy(GemmNPerThreadSubC) >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0,
"wrong!");
// output is a 10d tensor
constexpr index_t N1 = NPerBlock;
constexpr index_t W3 = GemmNPerThreadSubC / NPerBlock;
constexpr index_t W2 = GemmNLevel0Cluster * GemmNLevel1Cluster;
constexpr index_t W1 = WoPerBlock / f_dummy(W2 * W3);
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc = make_ConstantTensorDescriptor(
Sequence<K / (K1 * K2), K1, K2, Ho, Wo / (W1 * W2 * W3), W1, W2, W3, N / N1, N1>{});
constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread / K2, 1, K2, HoPerThread, 1, W1, 1, W3, 1, N1>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "out_10d_thread_desc");
print_ConstantTensorDescriptor(out_k_h_w_n_global_desc,
"out_k_h_w_n_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "out_10d_global_desc");
for(index_t i = 0; i < 64; ++i)
{
printf("out %f, ", p_out_thread[i]);
}
}
#endif
threadwise_tensor_slice_copy(out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_k_h_w_n_global_desc.GetOffsetFromMultiIndex(
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin,
n_block_data_begin + n_thread_data_begin),
out_10d_thread_desc.GetLengths(),
Number<OutThreadCopyDataPerWrite_N>{});
});
}
};
} // namespace ck
#endif
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r3_chwn_cyxk_khwn.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_CHWN_CYXK_KHWN
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_CHWN_CYXK_KHWN
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "threadwise_tensor_slice_copy.hpp"
#include "threadwise_4d_tensor_op.hpp"
#include "blockwise_batched_gemm.hpp"
namespace ck {
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
index_t NPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
index_t NPerThread,
index_t KPerThread,
index_t HoPerThread,
index_t WoPerThread,
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 InBlockCopyClusterLengths_CHWN,
index_t InBlockCopyDataPerRead_N,
index_t WeiBlockCopyDataPerRead_K,
index_t OutThreadCopyDataPerWrite_N>
struct GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn
{
__device__ void Run(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global) const
{
// be careful of this assertion
static_assert(
NPerBlock % NPerThread == 0 &&
((GemmNPerThreadSubC <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0) ||
(GemmNPerThreadSubC >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0)),
"wrong!");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_c_h_w_n_global_desc = InGlobalDesc{};
constexpr auto wei_c_y_x_k_global_desc = WeiGlobalDesc{};
constexpr auto out_k_h_w_n_global_desc = OutGlobalDesc{};
constexpr index_t C = in_c_h_w_n_global_desc.GetLength(I0);
constexpr index_t K = out_k_h_w_n_global_desc.GetLength(I0);
constexpr index_t Ho = out_k_h_w_n_global_desc.GetLength(I1);
constexpr index_t Wo = out_k_h_w_n_global_desc.GetLength(I2);
constexpr index_t N = out_k_h_w_n_global_desc.GetLength(I3);
constexpr index_t Y = wei_c_y_x_k_global_desc.GetLength(I1);
constexpr index_t X = wei_c_y_x_k_global_desc.GetLength(I2);
// divide block work: [K, Ho, Wo, N]
static_assert(N % NPerBlock == 0 && K % KPerBlock == 0 && C % CPerBlock == 0 &&
Ho % HoPerBlock == 0 && Wo % WoPerBlock == 0,
"wrong! cannot evenly divide work for workgroup ");
constexpr index_t NBlockWork = math::integer_divide_ceil(N, NPerBlock);
constexpr index_t KBlockWork = math::integer_divide_ceil(K, KPerBlock);
constexpr index_t HBlockWork = math::integer_divide_ceil(Ho, HoPerBlock);
constexpr index_t WBlockWork = math::integer_divide_ceil(Wo, WoPerBlock);
constexpr auto block_work_desc = make_ConstantTensorDescriptor(
Sequence<NBlockWork, KBlockWork, HBlockWork, WBlockWork>{});
const auto block_work_multi_id =
block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
const index_t n_block_data_begin = block_work_multi_id[0] * NPerBlock;
const index_t k_block_data_begin = block_work_multi_id[1] * KPerBlock;
const index_t ho_block_data_begin = block_work_multi_id[2] * HoPerBlock;
const index_t wo_block_data_begin = block_work_multi_id[3] * WoPerBlock;
const index_t hi_block_data_begin = ho_block_data_begin;
const index_t wi_block_data_begin = wo_block_data_begin;
// global tensor view
constexpr auto wei_c_k_global_desc =
make_ConstantTensorDescriptor(Sequence<C, K>{}, Sequence<Y * X * K, 1>{});
// LDS tensor view
// be careful of alignment
constexpr index_t max_align = math::lcm(InBlockCopyDataPerRead_N,
WeiBlockCopyDataPerRead_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, HoPerBlock, WoPerBlock, NPerBlock>{},
Number<InBlockCopyDataPerRead_N>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with alignment
static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not meet");
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
// tensor view of threadwise output in register
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
// blockwise copy
// input: format is [C, Hi, Wi, N]
const auto blockwise_in_copy =
Blockwise4dTensorCopy3<BlockSize,
Float,
decltype(in_c_h_w_n_global_desc),
decltype(in_c_h_w_n_block_desc),
decltype(in_c_h_w_n_block_desc.GetLengths()),
InBlockCopyClusterLengths_CHWN,
InBlockCopyDataPerRead_N>{};
// blockwise wei copy
// format is [CPerBlock, X * KPerBlock]
const auto blockwise_wei_copy =
Blockwise2dTensorCopy3<BlockSize,
Float,
decltype(wei_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopyDataPerRead_K>{};
// a series of blockwise batched GEMM
// C_matrix += transpose(A_matrix) * B_matrix
// A_matrix and B_matrix saved in LDS, C_matrix saved in register
// A_matrix[C,K] is a sub-matrix of wei_block[C,K]
// B_matrix[C,Wo*N] is a sub-matrix of in_block[C,Hi,Wi,N]
// C_matrix[K,Wo*N] is a sub-matrix of out_block[K,Ho,Wo,N]
constexpr auto a_c_k_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_c_k_block_desc.GetStride(I0)>{});
constexpr auto b_c_wn_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
Number<WoPerBlock * NPerBlock>{},
Number<in_c_h_w_n_block_desc.GetStride(I0)>{});
constexpr auto c_k_wn_thread_mtx_desc =
make_ConstantMatrixDescriptor(Number<KPerThread>{},
Number<WoPerThread * NPerThread>{},
Number<out_k_h_w_n_thread_desc.GetStride(I0)>{});
const auto blockwise_batch_gemm =
BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2<
BlockSize,
decltype(a_c_k_block_mtx_desc),
decltype(b_c_wn_block_mtx_desc),
decltype(c_k_wn_thread_mtx_desc),
0,
in_c_h_w_n_block_desc.GetStride(I1),
out_k_h_w_n_thread_desc.GetStride(I1),
HoPerBlock,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
HoPerThread,
GemmDataPerReadA,
GemmDataPerReadB>{};
// choose GEMM implementation here
const auto run_blockwise_batch_gemm = [&](auto... Xs) {
#if 0
return blockwise_batch_gemm.Run(Xs...);
#elif 0
return blockwise_batch_gemm.Run_amd_asm(Xs...);
#else
return blockwise_batch_gemm.Run_asm_v2(Xs...);
#endif
};
// LDS: be careful of alignment
// TODO:: need to properly implement tensor descriptor with alignment
constexpr index_t in_block_space =
in_c_h_w_n_block_desc.GetElementSpace(Number<max_align>{});
constexpr index_t wei_block_space = wei_c_k_block_desc.GetElementSpace(Number<max_align>{});
__shared__ Float p_in_block[in_block_space];
__shared__ Float p_wei_block[wei_block_space];
// register
// C++ lambda doesn't capture array, use pointer instead
Float p_out_thread_data[out_k_h_w_n_thread_desc.GetElementSpace()];
Float* const p_out_thread = p_out_thread_data;
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_c_h_w_n_global_desc, "in_c_h_w_n_global_desc");
print_ConstantTensorDescriptor(wei_c_y_x_k_global_desc, "wei_c_y_x_k_global_desc");
print_ConstantTensorDescriptor(in_c_h_w_n_block_desc, "in_c_h_w_n_block_desc");
print_ConstantTensorDescriptor(wei_c_x_k_block_desc, "wei_c_x_k_block_desc");
printf("in_block_space %u, wei_block_space %u\n", in_block_space, wei_block_space);
}
#endif
// set threadwise output tensor to 0
threadwise_4d_tensor_set_zero(out_k_h_w_n_thread_desc, p_out_thread);
#if 1
const Float* p_in_global_block_offset =
p_in_global +
in_c_h_w_n_global_desc.GetOffsetFromMultiIndex(
0, hi_block_data_begin, wi_block_data_begin, n_block_data_begin);
const Float* p_wei_global_block_offset =
p_wei_global +
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, 0, 0, k_block_data_begin);
for(index_t c_block_data_begin = 0; c_block_data_begin < C; c_block_data_begin += CPerBlock,
p_in_global_block_offset += CPerBlock * in_c_h_w_n_global_desc.GetStride(I0),
p_wei_global_block_offset += CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0))
{
for(index_t y = 0; y < Y; ++y)
{
#pragma unroll
for(index_t x = 0; x < X; ++x)
{
blockwise_in_copy.Run(
p_in_global_block_offset +
in_c_h_w_n_global_desc.GetOffsetFromMultiIndex(0, y, x, 0),
p_in_block);
blockwise_wei_copy.Run(
p_wei_global_block_offset +
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, x, 0),
p_wei_block);
__syncthreads();
run_blockwise_batch_gemm(p_wei_block, p_in_block, p_out_thread);
__syncthreads();
}
}
}
#else
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
const Float* p_in_global_block_offset =
p_in_global +
in_c_h_w_n_global_desc.GetOffsetFromMultiIndex(
0, hi_block_data_begin + y, wi_block_data_begin + x, n_block_data_begin);
const Float* p_wei_global_block_offset =
p_wei_global +
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, x, k_block_data_begin);
for(index_t c_block_data_begin = 0; c_block_data_begin < C;
c_block_data_begin += CPerBlock,
p_in_global_block_offset +=
CPerBlock * in_c_h_w_n_global_desc.GetStride(I0),
p_wei_global_block_offset +=
CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0))
{
blockwise_in_copy.Run(p_in_global_block_offset, p_in_block);
blockwise_wei_copy.Run(p_wei_global_block_offset, p_wei_block);
__syncthreads();
run_blockwise_batch_gemm(p_wei_block, p_in_block, p_out_thread);
__syncthreads();
}
}
}
#endif
// output: register to global mem,
const auto c_thread_mtx_begin =
blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t k_thread_data_begin = c_thread_mtx_begin.row;
const index_t ho_thread_data_begin = c_thread_mtx_begin.batch;
const index_t wo_thread_data_begin = c_thread_mtx_begin.col / NPerBlock;
const index_t n_thread_data_begin = c_thread_mtx_begin.col % NPerBlock;
static_if<GemmNPerThreadSubC <= NPerBlock>{}([&](auto f_dummy) { // f_dummy do nothing but
// perfect forwarding.
// Using this trick to
// make this lambda a generic lambda, so it won't be compiled until
// instantiated
static_assert(
(f_dummy(GemmNPerThreadSubC) <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0),
"wrong!");
// output is a 10d tensor
constexpr index_t N2 = GemmNPerThreadSubC;
constexpr index_t N1 = NPerBlock / N2;
constexpr index_t W2 =
(GemmNLevel0Cluster * GemmNLevel1Cluster) / f_dummy(NPerBlock / GemmNPerThreadSubC);
constexpr index_t W1 = WoPerBlock / W2;
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc =
make_ConstantTensorDescriptor(Sequence<K / (K1 * K2),
K1,
K2,
Ho,
Wo / (W1 * W2),
W1,
W2,
N / f_dummy(N1 * N2),
N1,
N2>{});
constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread / K2, 1, K2, HoPerThread, 1, W1, 1, 1, 1, N2>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "out_10d_thread_desc");
print_ConstantTensorDescriptor(out_k_h_w_n_global_desc,
"out_k_h_w_n_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "out_10d_global_desc");
}
#endif
threadwise_tensor_slice_copy(out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_k_h_w_n_global_desc.GetOffsetFromMultiIndex(
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin,
n_block_data_begin + n_thread_data_begin),
out_10d_thread_desc.GetLengths(),
Number<OutThreadCopyDataPerWrite_N>{});
}).Else([&](auto f_dummy) {
static_assert(f_dummy(GemmNPerThreadSubC) >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0,
"wrong!");
// output is a 10d tensor
constexpr index_t N1 = NPerBlock;
constexpr index_t W3 = GemmNPerThreadSubC / NPerBlock;
constexpr index_t W2 = GemmNLevel0Cluster * GemmNLevel1Cluster;
constexpr index_t W1 = WoPerBlock / f_dummy(W2 * W3);
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc = make_ConstantTensorDescriptor(
Sequence<K / (K1 * K2), K1, K2, Ho, Wo / (W1 * W2 * W3), W1, W2, W3, N / N1, N1>{});
constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread / K2, 1, K2, HoPerThread, 1, W1, 1, W3, 1, N1>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "out_10d_thread_desc");
print_ConstantTensorDescriptor(out_k_h_w_n_global_desc,
"out_k_h_w_n_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "out_10d_global_desc");
for(index_t i = 0; i < 64; ++i)
{
printf("out %f, ", p_out_thread[i]);
}
}
#endif
threadwise_tensor_slice_copy(out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_k_h_w_n_global_desc.GetOffsetFromMultiIndex(
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin,
n_block_data_begin + n_thread_data_begin),
out_10d_thread_desc.GetLengths(),
Number<OutThreadCopyDataPerWrite_N>{});
});
}
};
} // namespace ck
#endif
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r3_chwn_cyxk_khwn_lds_double_buffer.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_CHWN_CYXK_KHWN_LDS_DOUBLE_BUFFER
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_CHWN_CYXK_KHWN_LDS_DOUBLE_BUFFER
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "threadwise_tensor_slice_copy.hpp"
#include "threadwise_4d_tensor_op.hpp"
#include "blockwise_batched_gemm.hpp"
namespace ck {
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
index_t NPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
index_t NPerThread,
index_t KPerThread,
index_t HoPerThread,
index_t WoPerThread,
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 InBlockCopyClusterLengths_CHWN,
index_t InBlockCopyDataPerRead_N,
index_t WeiBlockCopyDataPerRead_K,
index_t OutThreadCopyDataPerWrite_N>
struct GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn_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
{
// be careful of this assertion
static_assert(
NPerBlock % NPerThread == 0 &&
((GemmNPerThreadSubC <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0) ||
(GemmNPerThreadSubC >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0)),
"wrong!");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_c_h_w_n_global_desc = InGlobalDesc{};
constexpr auto wei_c_y_x_k_global_desc = WeiGlobalDesc{};
constexpr auto out_k_h_w_n_global_desc = OutGlobalDesc{};
constexpr index_t C = in_c_h_w_n_global_desc.GetLength(I0);
constexpr index_t K = out_k_h_w_n_global_desc.GetLength(I0);
constexpr index_t Ho = out_k_h_w_n_global_desc.GetLength(I1);
constexpr index_t Wo = out_k_h_w_n_global_desc.GetLength(I2);
constexpr index_t N = out_k_h_w_n_global_desc.GetLength(I3);
constexpr index_t Y = wei_c_y_x_k_global_desc.GetLength(I1);
constexpr index_t X = wei_c_y_x_k_global_desc.GetLength(I2);
constexpr index_t HiPerBlock = HoPerBlock + Y - 1;
constexpr index_t WiPerBlock = WoPerBlock + X - 1;
// assert for LDS double buffer
static_assert(C % (2 * CPerBlock) == 0, "C cannot be evenly divided");
// divide block work: [K, Ho, Wo, N]
static_assert(N % NPerBlock == 0 && K % KPerBlock == 0 && C % CPerBlock == 0 &&
Ho % HoPerBlock == 0 && Wo % WoPerBlock == 0,
"wrong! cannot evenly divide work for workgroup ");
constexpr index_t KBlockWork = math::integer_divide_ceil(K, KPerBlock);
constexpr index_t HBlockWork = math::integer_divide_ceil(Ho, HoPerBlock);
constexpr index_t WBlockWork = math::integer_divide_ceil(Wo, WoPerBlock);
constexpr index_t NBlockWork = math::integer_divide_ceil(N, NPerBlock);
constexpr auto block_work_desc = make_ConstantTensorDescriptor_packed(
Sequence<KBlockWork, HBlockWork, WBlockWork, NBlockWork>{});
const auto block_work_multi_id =
block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
const index_t k_block_data_begin = block_work_multi_id[0] * KPerBlock;
const index_t ho_block_data_begin = block_work_multi_id[1] * HoPerBlock;
const index_t wo_block_data_begin = block_work_multi_id[2] * WoPerBlock;
const index_t n_block_data_begin = block_work_multi_id[3] * NPerBlock;
const index_t hi_block_data_begin = ho_block_data_begin;
const index_t wi_block_data_begin = wo_block_data_begin;
// global tensor view
constexpr auto wei_c_k_global_desc = wei_c_y_x_k_global_desc.Extract(I0, I3);
// LDS tensor view
// be careful of alignment
constexpr index_t max_align = math::lcm(InBlockCopyDataPerRead_N,
WeiBlockCopyDataPerRead_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, HoPerBlock, WoPerBlock, NPerBlock>{},
Number<InBlockCopyDataPerRead_N>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with alignment
static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not meet");
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
// tensor view of threadwise output in register
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_packed(
Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
// blockwise copy
// input: format is [C, Hi, Wi, N]
#if 0
const auto blockwise_in_copy =
Blockwise4dTensorCopy1<BlockSize,
Float,
decltype(in_c_h_w_n_global_desc),
decltype(in_c_h_w_n_block_desc),
decltype(in_c_h_w_n_block_desc.GetLengths()),
InBlockCopyDataPerRead_N>{};
#else
const auto blockwise_in_copy =
Blockwise4dTensorCopy3<BlockSize,
Float,
decltype(in_c_h_w_n_global_desc),
decltype(in_c_h_w_n_block_desc),
decltype(in_c_h_w_n_block_desc.GetLengths()),
InBlockCopyClusterLengths_CHWN,
InBlockCopyDataPerRead_N>{};
#endif
// blockwise wei copy
// format is [CPerBlock, X * KPerBlock]
const auto blockwise_wei_copy =
Blockwise2dTensorCopy3<BlockSize,
Float,
decltype(wei_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopyDataPerRead_K>({0, 0}, {0, 0});
// a series of blockwise batched GEMM
// C_matrix += transpose(A_matrix) * B_matrix
// A_matrix and B_matrix saved in LDS, C_matrix saved in register
// A_matrix[C,K] is a sub-matrix of wei_block[C,K]
// B_matrix[C,Wo*N] is a sub-matrix of in_block[C,Hi,Wi,N]
// C_matrix[K,Wo*N] is a sub-matrix of out_block[K,Ho,Wo,N]
constexpr auto a_c_k_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_c_k_block_desc.GetStride(I0)>{});
constexpr auto b_c_wn_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
Number<WoPerBlock * NPerBlock>{},
Number<in_c_h_w_n_block_desc.GetStride(I0)>{});
constexpr auto c_k_wn_thread_mtx_desc =
make_ConstantMatrixDescriptor(Number<KPerThread>{},
Number<WoPerThread * NPerThread>{},
Number<out_k_h_w_n_thread_desc.GetStride(I0)>{});
const auto blockwise_batch_gemm =
BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2<
BlockSize,
decltype(a_c_k_block_mtx_desc),
decltype(b_c_wn_block_mtx_desc),
decltype(c_k_wn_thread_mtx_desc),
0,
in_c_h_w_n_block_desc.GetStride(I1),
out_k_h_w_n_thread_desc.GetStride(I1),
HoPerBlock,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
HoPerThread,
GemmDataPerReadA,
GemmDataPerReadB>{};
// choose GEMM implementation here
const auto run_blockwise_batch_gemm = [&](auto... Xs) {
#if 1
return blockwise_batch_gemm.Run(Xs...);
#elif 0
return blockwise_batch_gemm.Run_amd_asm(Xs...);
#else
return blockwise_batch_gemm.Run_asm_v2(Xs...);
#endif
};
// LDS: be careful of alignment
constexpr index_t in_block_space =
in_c_h_w_n_block_desc.GetElementSpace(Number<max_align>{});
constexpr index_t wei_block_space = wei_c_k_block_desc.GetElementSpace(Number<max_align>{});
// LDS double buffer
__shared__ Float p_in_block_double[2 * in_block_space];
__shared__ Float p_wei_block_double[2 * wei_block_space];
// register
// C++ lambda doesn't capture array, use pointer instead
Float p_out_thread_data[out_k_h_w_n_thread_desc.GetElementSpace()];
Float* const p_out_thread = p_out_thread_data;
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_c_h_w_n_global_desc, "in_c_h_w_n_global_desc");
print_ConstantTensorDescriptor(wei_c_y_x_k_global_desc, "wei_c_y_x_k_global_desc");
print_ConstantTensorDescriptor(in_c_h_w_n_block_desc, "in_c_h_w_n_block_desc");
print_ConstantTensorDescriptor(wei_c_x_k_block_desc, "wei_c_x_k_block_desc");
printf("in_block_space %u, wei_block_space %u\n", in_block_space, wei_block_space);
}
#endif
// set threadwise output to 0
threadwise_matrix_set_zero(c_k_wn_thread_mtx_desc, p_out_thread);
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
const Float* p_in_global_block_offset =
p_in_global +
in_c_h_w_n_global_desc.GetOffsetFromMultiIndex(
0, hi_block_data_begin + y, wi_block_data_begin + x, n_block_data_begin);
const Float* p_wei_global_block_offset =
p_wei_global +
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, x, k_block_data_begin);
// LDS double buffer: preload data into LDS
{
Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global_block_offset,
p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
p_wei_register_clipboard);
blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard,
p_in_block_double);
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard,
p_wei_block_double);
}
// LDS double buffer: main body
for(index_t c_block_data_begin = 0; c_block_data_begin + 2 * CPerBlock < C;
c_block_data_begin += 2 * CPerBlock)
{
#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()];
p_in_global_block_offset +=
CPerBlock * in_c_h_w_n_global_desc.GetStride(I0);
p_wei_global_block_offset +=
CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global_block_offset,
p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
p_wei_register_clipboard);
run_blockwise_batch_gemm(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
p_in_global_block_offset += CPerBlock * in_c_h_w_n_global_desc.GetStride(I0);
p_wei_global_block_offset += CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global_block_offset,
p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
p_wei_register_clipboard);
// LDS double buffer: GEMM on current data
run_blockwise_batch_gemm(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
run_blockwise_batch_gemm(p_wei_block_double + wei_block_space,
p_in_block_double + in_block_space,
p_out_thread);
}
}
}
// output: register to global mem,
const auto c_thread_mtx_begin =
blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t k_thread_data_begin = c_thread_mtx_begin.row;
const index_t ho_thread_data_begin = c_thread_mtx_begin.batch;
const index_t wo_thread_data_begin = c_thread_mtx_begin.col / NPerBlock;
const index_t n_thread_data_begin = c_thread_mtx_begin.col % NPerBlock;
static_if<GemmNPerThreadSubC <= NPerBlock>{}([&](auto fwd) {
// fwd do nothing but perfect forwarding.
// Using this trick to make this lambda a generic lambda, so it won't be compiled until
// being instantiated here
static_assert(
(fwd(GemmNPerThreadSubC) <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0),
"wrong!");
// output is a 10d tensor
constexpr index_t N2 = GemmNPerThreadSubC;
constexpr index_t N1 = NPerBlock / N2;
constexpr index_t W2 =
(GemmNLevel0Cluster * GemmNLevel1Cluster) / fwd(NPerBlock / GemmNPerThreadSubC);
constexpr index_t W1 = WoPerBlock / W2;
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc = fwd(out_k_h_w_n_global_desc)
.Fold(I3, Number<N1>{}, Number<N2>{})
.Fold(I2, Number<W1>{}, Number<W2>{})
.Fold(I0, Number<K1>{}, Number<K2>{});
constexpr auto out_10d_thread_desc = fwd(out_k_h_w_n_thread_desc)
.Fold(I3, Number<1>{}, Number<N2>{})
.Fold(I2, Number<W1>{}, Number<1>{})
.Fold(I0, Number<1>{}, Number<K2>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"a: out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "a: out_10d_thread_desc");
print_ConstantTensorDescriptor(out_k_h_w_n_global_desc,
"a: out_k_h_w_n_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "a: out_10d_global_desc");
}
#endif
threadwise_tensor_slice_copy(out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_k_h_w_n_global_desc.GetOffsetFromMultiIndex(
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin,
n_block_data_begin + n_thread_data_begin),
out_10d_thread_desc.GetLengths(),
Number<OutThreadCopyDataPerWrite_N>{});
}).Else([&](auto fwd) {
static_assert(fwd(GemmNPerThreadSubC) >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0,
"wrong!");
// output is a 10d tensor
constexpr index_t N1 = NPerBlock;
constexpr index_t W3 = GemmNPerThreadSubC / NPerBlock;
constexpr index_t W2 = GemmNLevel0Cluster * GemmNLevel1Cluster;
constexpr index_t W1 = WoPerBlock / fwd(W2 * W3);
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc =
fwd(out_k_h_w_n_global_desc)
.Fold(I3, Number<N1>{})
.Fold(I2, Number<W1>{}, Number<W2>{}, Number<W3>{})
.Fold(I0, Number<K1>{}, Number<K2>{});
constexpr auto out_10d_thread_desc =
fwd(out_k_h_w_n_thread_desc)
.Fold(I3, Number<N1>{})
.Fold(I2, Number<W1>{}, Number<1>{}, Number<W3>{})
.Fold(I0, Number<1>{}, Number<K2>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"b: out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "b: out_10d_thread_desc");
print_ConstantTensorDescriptor(out_k_h_w_n_global_desc,
"b: out_k_h_w_n_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "b: out_10d_global_desc");
}
#endif
threadwise_tensor_slice_copy(out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_k_h_w_n_global_desc.GetOffsetFromMultiIndex(
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin,
n_block_data_begin + n_thread_data_begin),
out_10d_thread_desc.GetLengths(),
Number<OutThreadCopyDataPerWrite_N>{});
});
}
};
} // namespace ck
#endif
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_nkhw.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_NCHW_CYXK_NKHW
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_NCHW_CYXK_NKHW
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_tensor_slice_copy.hpp"
#include "threadwise_tensor_slice_copy.hpp"
#include "threadwise_generic_tensor_op.hpp"
#include "blockwise_batched_gemm.hpp"
namespace ck {
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
index_t NPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
index_t NPerThread,
index_t KPerThread,
index_t HoPerThread,
index_t WoPerThread,
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 InBlockReorderSrcSubLengths_NCHW,
class InBlockReorderSrcClusterLengths_NCHW,
class InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW,
index_t InBlockReorderDataPerRead_W,
index_t InBlockReorderDataPerWrite_N,
class WeiBlockCopyClusterLengths_CK, // not used
index_t WeiBlockCopyDataPerRead_K,
index_t OutThreadCopyDataPerWrite_W>
struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_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
{
// be careful of this assertion
static_assert(
NPerBlock % NPerThread == 0 &&
((GemmNPerThreadSubC <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0) ||
(GemmNPerThreadSubC >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0)),
"wrong!");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_n_c_h_w_global_desc = InGlobalDesc{};
constexpr auto wei_c_y_x_k_global_desc = WeiGlobalDesc{};
constexpr auto out_n_k_h_w_global_desc = OutGlobalDesc{};
constexpr index_t C = in_n_c_h_w_global_desc.GetLength(I1);
constexpr index_t N = out_n_k_h_w_global_desc.GetLength(I0);
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_c_y_x_k_global_desc.GetLength(I1);
constexpr index_t X = wei_c_y_x_k_global_desc.GetLength(I2);
// divide block work: [N, K, Ho, Wo]
static_assert(N % NPerBlock == 0 && K % KPerBlock == 0 && C % CPerBlock == 0 &&
Ho % HoPerBlock == 0 && Wo % WoPerBlock == 0,
"wrong! cannot evenly divide work for workgroup ");
constexpr index_t NBlockWork = math::integer_divide_ceil(N, NPerBlock);
constexpr index_t KBlockWork = math::integer_divide_ceil(K, KPerBlock);
constexpr index_t HBlockWork = math::integer_divide_ceil(Ho, HoPerBlock);
constexpr index_t WBlockWork = math::integer_divide_ceil(Wo, WoPerBlock);
constexpr auto block_work_desc = make_ConstantTensorDescriptor_packed(
Sequence<NBlockWork, KBlockWork, HBlockWork, WBlockWork>{});
const auto block_work_multi_id =
block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
const index_t n_block_data_begin = block_work_multi_id[0] * NPerBlock;
const index_t k_block_data_begin = block_work_multi_id[1] * KPerBlock;
const index_t ho_block_data_begin = block_work_multi_id[2] * HoPerBlock;
const index_t wo_block_data_begin = block_work_multi_id[3] * WoPerBlock;
const index_t hi_block_data_begin = ho_block_data_begin;
const index_t wi_block_data_begin = wo_block_data_begin;
// global tensor view
constexpr auto wei_c_k_global_desc =
make_ConstantTensorDescriptor(Sequence<C, K>{}, Sequence<Y * X * K, 1>{});
// LDS tensor view
// be careful of alignment
constexpr index_t max_align = math::lcm(InBlockReorderDataPerWrite_N,
WeiBlockCopyDataPerRead_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, HoPerBlock, WoPerBlock, NPerBlock>{},
Number<InBlockReorderDataPerWrite_N>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with alignment
static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not meet");
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
// tensor view of threadwise output in register
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_packed(
Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
// blockwise copy
// input: format is [N, C, Hi, Wi] to [C, Hi, Wi, N]
constexpr auto map_chwn2nchw = Sequence<1, 2, 3, 0>{};
const auto blockwise_in_copy_reorder = BlockwiseTensorSliceReorderCopy_v3<
BlockSize,
Float,
decltype(in_n_c_h_w_global_desc),
decltype(in_c_h_w_n_block_desc),
Sequence<NPerBlock, CPerBlock, HoPerBlock, WoPerBlock>,
InBlockReorderSrcSubLengths_NCHW,
InBlockReorderSrcClusterLengths_NCHW,
decltype(map_chwn2nchw),
InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW,
InBlockReorderDataPerRead_W,
InBlockReorderDataPerWrite_N>({0, 0, 0, 0}, {0, 0, 0, 0});
// blockwise wei copy
// format is [CPerBlock, KPerBlock]
const auto blockwise_wei_copy =
Blockwise2dTensorCopy3<BlockSize,
Float,
decltype(wei_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopyDataPerRead_K>({0, 0}, {0, 0});
// a series of blockwise batched GEMM
// C_matrix += transpose(A_matrix) * B_matrix
// A_matrix and B_matrix saved in LDS, C_matrix saved in register
// A_matrix[C,K] is a sub-matrix of wei_block[C,K]
// B_matrix[C,Wo*N] is a sub-matrix of in_block[C,Hi,Wi,N]
// C_matrix[K,Wo*N] is a sub-matrix of out_block[K,Ho,Wo,N]
constexpr auto a_c_k_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_c_k_block_desc.GetStride(I0)>{});
constexpr auto b_c_wn_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
Number<WoPerBlock * NPerBlock>{},
Number<in_c_h_w_n_block_desc.GetStride(I0)>{});
constexpr auto c_k_wn_thread_mtx_desc =
make_ConstantMatrixDescriptor(Number<KPerThread>{},
Number<WoPerThread * NPerThread>{},
Number<out_k_h_w_n_thread_desc.GetStride(I0)>{});
const auto blockwise_batch_gemm =
BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2<
BlockSize,
decltype(a_c_k_block_mtx_desc),
decltype(b_c_wn_block_mtx_desc),
decltype(c_k_wn_thread_mtx_desc),
0,
in_c_h_w_n_block_desc.GetStride(I1),
out_k_h_w_n_thread_desc.GetStride(I1),
HoPerBlock,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
HoPerThread,
GemmDataPerReadA,
GemmDataPerReadB>{};
// choose GEMM implementation here
const auto run_blockwise_batch_gemm = [&](auto... Xs) {
#if 1
return blockwise_batch_gemm.Run(Xs...);
#elif 0
return blockwise_batch_gemm.Run_amd_asm(Xs...);
#else
return blockwise_batch_gemm.Run_asm_v2(Xs...);
#endif
};
// LDS: be careful of alignment
constexpr index_t in_block_space =
in_c_h_w_n_block_desc.GetElementSpace(Number<max_align>{});
constexpr index_t wei_block_space = wei_c_k_block_desc.GetElementSpace(Number<max_align>{});
__shared__ Float p_in_block[in_block_space];
__shared__ Float p_wei_block[wei_block_space];
// register
// C++ lambda doesn't capture array, use pointer instead
Float p_out_thread_data[out_k_h_w_n_thread_desc.GetElementSpace()];
Float* const p_out_thread = p_out_thread_data;
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_c_h_w_n_global_desc, "in_c_h_w_n_global_desc");
print_ConstantTensorDescriptor(wei_c_y_x_k_global_desc, "wei_c_y_x_k_global_desc");
print_ConstantTensorDescriptor(in_c_h_w_n_block_desc, "in_c_h_w_n_block_desc");
print_ConstantTensorDescriptor(wei_c_k_block_desc, "wei_c_k_block_desc");
printf("in_block_space %u, wei_block_space %u\n", in_block_space, wei_block_space);
}
#endif
// set threadwise output tensor to 0
threadwise_generic_tensor_set_zero(out_k_h_w_n_thread_desc, p_out_thread);
#if 0
const Float* p_in_global_block_offset =
p_in_global +
in_n_c_h_w_global_desc.GetOffsetFromMultiIndex(
n_block_data_begin, 0, hi_block_data_begin, wi_block_data_begin);
const Float* p_wei_global_block_offset =
p_wei_global + wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, 0, 0, k_block_data_begin);
for(index_t c_block_data_begin = 0; c_block_data_begin < C; c_block_data_begin += CPerBlock,
p_in_global_block_offset += CPerBlock * in_n_c_h_w_global_desc.GetStride(I1),
p_wei_global_block_offset += CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0))
{
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
blockwise_in_copy_reorder.Run(p_in_global_block_offset +
in_n_c_h_w_global_desc.GetOffsetFromMultiIndex(0, 0, y, x),
p_in_block);
blockwise_wei_copy.Run(p_wei_global_block_offset +
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, x, 0),
p_wei_block);
__syncthreads();
run_blockwise_batch_gemm(p_wei_block, p_in_block, p_out_thread);
__syncthreads();
}
}
}
#else
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
const Float* p_in_global_block_offset =
p_in_global +
in_n_c_h_w_global_desc.GetOffsetFromMultiIndex(
n_block_data_begin, 0, hi_block_data_begin + y, wi_block_data_begin + x);
const Float* p_wei_global_block_offset =
p_wei_global +
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, x, k_block_data_begin);
for(index_t c_block_data_begin = 0; c_block_data_begin < C;
c_block_data_begin += CPerBlock,
p_in_global_block_offset +=
CPerBlock * in_n_c_h_w_global_desc.GetStride(I1),
p_wei_global_block_offset +=
CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0))
{
blockwise_in_copy_reorder.Run(p_in_global_block_offset, p_in_block);
blockwise_wei_copy.Run(p_wei_global_block_offset, p_wei_block);
__syncthreads();
run_blockwise_batch_gemm(p_wei_block, p_in_block, p_out_thread);
__syncthreads();
}
}
}
#endif
// output: register to global mem,
const auto c_thread_mtx_begin =
blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t k_thread_data_begin = c_thread_mtx_begin.row;
const index_t ho_thread_data_begin = c_thread_mtx_begin.batch;
const index_t wo_thread_data_begin = c_thread_mtx_begin.col / NPerBlock;
const index_t n_thread_data_begin = c_thread_mtx_begin.col % NPerBlock;
static_if<GemmNPerThreadSubC <= NPerBlock>{}([&](auto fwd) {
// fwd do nothing but perfect forwarding.
// Using this trick to make this lambda a generic lambda, so it won't be compiled until
// begin instantiated here
static_assert(
(fwd(GemmNPerThreadSubC) <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0),
"wrong!");
// output is a 10d tensor
constexpr index_t N2 = GemmNPerThreadSubC;
constexpr index_t N1 = NPerBlock / N2;
constexpr index_t W2 =
(GemmNLevel0Cluster * GemmNLevel1Cluster) / fwd(NPerBlock / GemmNPerThreadSubC);
constexpr index_t W1 = WoPerBlock / W2;
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc = fwd(out_n_k_h_w_global_desc)
.Fold(I3, Number<W1>{}, Number<W2>{})
.Fold(I1, Number<K1>{}, Number<K2>{})
.Fold(I0, Number<N1>{}, Number<N2>{});
constexpr auto out_10d_thread_desc = fwd(out_k_h_w_n_thread_desc)
.Fold(I3, Number<1>{}, Number<N2>{})
.Fold(I2, Number<W1>{}, Number<1>{})
.Fold(I0, Number<1>{}, Number<K2>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"a: out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "a: out_10d_thread_desc");
print_ConstantTensorDescriptor(out_n_k_h_w_global_desc,
"a: out_n_k_h_w_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "a: out_10d_global_desc");
}
#endif
constexpr auto map_out_global2thread = Sequence<7, 8, 9, 0, 1, 2, 3, 4, 5, 6>{};
threadwise_tensor_slice_copy_reorder_given_dst2src_v2(
out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_n_k_h_w_global_desc.GetOffsetFromMultiIndex(
n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
out_10d_thread_desc.GetLengths(),
map_out_global2thread);
// Number<OutThreadCopyDataPerWrite_W>{});
}).Else([&](auto fwd) {
static_assert(fwd(GemmNPerThreadSubC) >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0,
"wrong!");
// output is a 10d tensor
constexpr index_t N1 = NPerBlock;
constexpr index_t W3 = GemmNPerThreadSubC / NPerBlock;
constexpr index_t W2 = GemmNLevel0Cluster * GemmNLevel1Cluster;
constexpr index_t W1 = WoPerBlock / fwd(W2 * W3);
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc =
fwd(out_n_k_h_w_global_desc)
.Fold(I3, Number<W1>{}, Number<W2>{}, Number<W3>{})
.Fold(I1, Number<K1>{}, Number<K2>{})
.Fold(I0, Number<N1>{});
constexpr auto out_10d_thread_desc =
fwd(out_k_h_w_n_thread_desc)
.Fold(I3, Number<N1>{})
.Fold(I2, Number<W1>{}, Number<1>{}, Number<W3>{})
.Fold(I0, Number<1>{}, Number<K2>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"b: out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "b: out_10d_thread_desc");
print_ConstantTensorDescriptor(out_n_k_h_w_global_desc,
"b: out_n_k_h_w_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "b: out_10d_global_desc");
}
#endif
constexpr auto map_out_global2thread = Sequence<8, 9, 0, 1, 2, 3, 4, 5, 6, 7>{};
#if 0
threadwise_tensor_slice_copy_reorder_given_dst2src_v3(
out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_n_k_h_w_global_desc.GetOffsetFromMultiIndex(
n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
out_10d_thread_desc.GetLengths(),
map_out_global2thread,
Number<OutThreadCopyDataPerWrite_W>{});
#else
threadwise_generic_tensor_slice_copy_v1(
out_10d_thread_desc.ReorderGivenNew2Old(map_out_global2thread),
p_out_thread,
make_zero_array<index_t, 10>(),
out_10d_global_desc,
p_out_global +
out_n_k_h_w_global_desc.GetOffsetFromMultiIndex(
n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
make_zero_array<index_t, 10>(),
out_10d_thread_desc.GetLengths().ReorderGivenNew2Old(map_out_global2thread),
arithmetic_sequence_gen<0, 10, 1>::type{},
Number<1>{});
#endif
});
}
};
} // namespace ck
#endif
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_nkhw_lds_double_buffer.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_NCHW_CYXK_NKHW_LDS_DOUBLE_BUFFER
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_NCHW_CYXK_NKHW_LDS_DOUBLE_BUFFER
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_tensor_slice_copy.hpp"
#include "threadwise_tensor_slice_copy.hpp"
#include "threadwise_generic_tensor_op.hpp"
#include "blockwise_batched_gemm.hpp"
namespace ck {
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
index_t NPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
index_t NPerThread,
index_t KPerThread,
index_t HoPerThread,
index_t WoPerThread,
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 InBlockReorderSrcSubLengths_NCHW,
class InBlockReorderSrcClusterLengths_NCHW,
class InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW,
index_t InBlockReorderDataPerRead_W,
index_t InBlockReorderDataPerWrite_N,
class WeiBlockCopyClusterLengths_CK, // not used
index_t WeiBlockCopyDataPerRead_K,
index_t OutThreadCopyDataPerWrite_W>
struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_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
{
// be careful of this assertion
static_assert(
NPerBlock % NPerThread == 0 &&
((GemmNPerThreadSubC <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0) ||
(GemmNPerThreadSubC >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0)),
"wrong!");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_n_c_h_w_global_desc = InGlobalDesc{};
constexpr auto wei_c_y_x_k_global_desc = WeiGlobalDesc{};
constexpr auto out_n_k_h_w_global_desc = OutGlobalDesc{};
constexpr index_t C = in_n_c_h_w_global_desc.GetLength(I1);
constexpr index_t N = out_n_k_h_w_global_desc.GetLength(I0);
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_c_y_x_k_global_desc.GetLength(I1);
constexpr index_t X = wei_c_y_x_k_global_desc.GetLength(I2);
// assert for LDS double buffer
static_assert(C % (2 * CPerBlock) == 0, "C cannot be evenly divided");
// divide block work: [K, Ho, Wo, N]
static_assert(N % NPerBlock == 0 && K % KPerBlock == 0 && C % CPerBlock == 0 &&
Ho % HoPerBlock == 0 && Wo % WoPerBlock == 0,
"wrong! cannot evenly divide work for workgroup ");
constexpr index_t NBlockWork = math::integer_divide_ceil(N, NPerBlock);
constexpr index_t KBlockWork = math::integer_divide_ceil(K, KPerBlock);
constexpr index_t HBlockWork = math::integer_divide_ceil(Ho, HoPerBlock);
constexpr index_t WBlockWork = math::integer_divide_ceil(Wo, WoPerBlock);
constexpr auto block_work_desc = make_ConstantTensorDescriptor_packed(
Sequence<NBlockWork, KBlockWork, HBlockWork, WBlockWork>{});
const auto block_work_multi_id =
block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
const index_t n_block_data_begin = block_work_multi_id[0] * NPerBlock;
const index_t k_block_data_begin = block_work_multi_id[1] * KPerBlock;
const index_t ho_block_data_begin = block_work_multi_id[2] * HoPerBlock;
const index_t wo_block_data_begin = block_work_multi_id[3] * WoPerBlock;
const index_t hi_block_data_begin = ho_block_data_begin;
const index_t wi_block_data_begin = wo_block_data_begin;
// global tensor view
constexpr auto wei_c_k_global_desc = wei_c_y_x_k_global_desc.Extract(I0, I3);
// LDS tensor view
// be careful of alignment
constexpr index_t max_align = math::lcm(InBlockReorderDataPerWrite_N,
WeiBlockCopyDataPerRead_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, HoPerBlock, WoPerBlock, NPerBlock>{},
Number<InBlockReorderDataPerWrite_N>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with multiple alignment requirements
static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not meet");
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
// tensor view of threadwise output in register
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_packed(
Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
// blockwise copy
// input: format is [N, C, Hi, Wi] to [C, Hi, Wi, N]
constexpr auto map_chwn2nchw = Sequence<1, 2, 3, 0>{};
const auto blockwise_in_copy_reorder = BlockwiseTensorSliceReorderCopy_v3<
BlockSize,
Float,
decltype(in_n_c_h_w_global_desc),
decltype(in_c_h_w_n_block_desc),
Sequence<NPerBlock, CPerBlock, HoPerBlock, WoPerBlock>,
InBlockReorderSrcSubLengths_NCHW,
InBlockReorderSrcClusterLengths_NCHW,
decltype(map_chwn2nchw),
InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW,
InBlockReorderDataPerRead_W,
InBlockReorderDataPerWrite_N>({0, 0, 0, 0}, {0, 0, 0, 0});
// blockwise wei copy
// format is [CPerBlock, KPerBlock]
const auto blockwise_wei_copy =
Blockwise2dTensorCopy3<BlockSize,
Float,
decltype(wei_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopyDataPerRead_K>({0, 0}, {0, 0});
// a series of blockwise batched GEMM
// C_matrix += transpose(A_matrix) * B_matrix
// A_matrix and B_matrix saved in LDS, C_matrix saved in register
// A_matrix[C,K] is a sub-matrix of wei_block[C,K]
// B_matrix[C,Wo*N] is a sub-matrix of in_block[C,Hi,Wi,N]
// C_matrix[K,Wo*N] is a sub-matrix of out_block[K,Ho,Wo,N]
constexpr auto a_c_k_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_c_k_block_desc.GetStride(I0)>{});
constexpr auto b_c_wn_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
Number<WoPerBlock * NPerBlock>{},
Number<in_c_h_w_n_block_desc.GetStride(I0)>{});
constexpr auto c_k_wn_thread_mtx_desc =
make_ConstantMatrixDescriptor(Number<KPerThread>{},
Number<WoPerThread * NPerThread>{},
Number<out_k_h_w_n_thread_desc.GetStride(I0)>{});
const auto blockwise_batch_gemm =
BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2<
BlockSize,
decltype(a_c_k_block_mtx_desc),
decltype(b_c_wn_block_mtx_desc),
decltype(c_k_wn_thread_mtx_desc),
0,
in_c_h_w_n_block_desc.GetStride(I1),
out_k_h_w_n_thread_desc.GetStride(I1),
HoPerBlock,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
HoPerThread,
GemmDataPerReadA,
GemmDataPerReadB>{};
// choose GEMM implementation here
const auto run_blockwise_batch_gemm = [&](auto... Xs) {
#if 1
return blockwise_batch_gemm.Run(Xs...);
#elif 0
return blockwise_batch_gemm.Run_amd_asm(Xs...);
#else
return blockwise_batch_gemm.Run_asm_v2(Xs...);
#endif
};
// LDS: be careful of alignment
constexpr index_t in_block_space =
in_c_h_w_n_block_desc.GetElementSpace(Number<max_align>{});
constexpr index_t wei_block_space = wei_c_k_block_desc.GetElementSpace(Number<max_align>{});
// LDS double buffer
__shared__ Float p_in_block_double[2 * in_block_space];
__shared__ Float p_wei_block_double[2 * wei_block_space];
// register
// C++ lambda doesn't capture array, use pointer instead
Float p_out_thread_data[out_k_h_w_n_thread_desc.GetElementSpace()];
Float* const p_out_thread = p_out_thread_data;
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_c_h_w_n_global_desc, "in_c_h_w_n_global_desc");
print_ConstantTensorDescriptor(wei_c_y_x_k_global_desc, "wei_c_y_x_k_global_desc");
print_ConstantTensorDescriptor(in_c_h_w_n_block_desc, "in_c_h_w_n_block_desc");
print_ConstantTensorDescriptor(wei_c_k_block_desc, "wei_c_k_block_desc");
printf("in_block_space %u, wei_block_space %u\n", in_block_space, wei_block_space);
}
#endif
// set threadwise output tensor to 0
threadwise_generic_tensor_set_zero(out_k_h_w_n_thread_desc, p_out_thread);
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
const Float* p_in_global_block_offset =
p_in_global +
in_n_c_h_w_global_desc.GetOffsetFromMultiIndex(
n_block_data_begin, 0, hi_block_data_begin + y, wi_block_data_begin + x);
const Float* p_wei_global_block_offset =
p_wei_global +
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, x, k_block_data_begin);
// LDS double buffer: preload data into LDS
{
Float p_in_register_clipboard[blockwise_in_copy_reorder
.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
blockwise_in_copy_reorder.RunLoadRegisterClipboard(p_in_global_block_offset,
p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
p_wei_register_clipboard);
blockwise_in_copy_reorder.RunStoreRegisterClipboard(p_in_register_clipboard,
p_in_block_double);
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard,
p_wei_block_double);
}
// LDS double buffer: main body
for(index_t c_block_data_begin = 0; c_block_data_begin + 2 * CPerBlock < C;
c_block_data_begin += 2 * CPerBlock)
{
#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_reorder
.GetRegisterClipboardSize()];
Float
p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
p_in_global_block_offset +=
CPerBlock * in_n_c_h_w_global_desc.GetStride(I1);
p_wei_global_block_offset +=
CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy_reorder.RunLoadRegisterClipboard(p_in_global_block_offset,
p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
p_wei_register_clipboard);
// LDS double buffer: GEMM on current data
run_blockwise_batch_gemm(p_wei_block_now, p_in_block_now, p_out_thread);
// LDS double buffer: store next data to LDS
blockwise_in_copy_reorder.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_reorder
.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
// even iteration
p_in_global_block_offset += CPerBlock * in_n_c_h_w_global_desc.GetStride(I1);
p_wei_global_block_offset += CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy_reorder.RunLoadRegisterClipboard(p_in_global_block_offset,
p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
p_wei_register_clipboard);
// LDS double buffer: GEMM on current data
run_blockwise_batch_gemm(p_wei_block_double, p_in_block_double, p_out_thread);
// LDS double buffer: store next data to LDS
blockwise_in_copy_reorder.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
run_blockwise_batch_gemm(p_wei_block_double + wei_block_space,
p_in_block_double + in_block_space,
p_out_thread);
}
}
}
// output: register to global mem,
const auto c_thread_mtx_begin =
blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t k_thread_data_begin = c_thread_mtx_begin.row;
const index_t ho_thread_data_begin = c_thread_mtx_begin.batch;
const index_t wo_thread_data_begin = c_thread_mtx_begin.col / NPerBlock;
const index_t n_thread_data_begin = c_thread_mtx_begin.col % NPerBlock;
static_if<GemmNPerThreadSubC <= NPerBlock>{}([&](auto fwd) {
// fwd do nothing but perfect forwarding.
// Using this trick to make this lambda a generic lambda, so it won't be compiled until
// begin instantiated here
static_assert(
(fwd(GemmNPerThreadSubC) <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0),
"wrong!");
// output is a 10d tensor
constexpr index_t N2 = GemmNPerThreadSubC;
constexpr index_t N1 = NPerBlock / N2;
constexpr index_t W2 =
(GemmNLevel0Cluster * GemmNLevel1Cluster) / fwd(NPerBlock / GemmNPerThreadSubC);
constexpr index_t W1 = WoPerBlock / W2;
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc = fwd(out_n_k_h_w_global_desc)
.Fold(I3, Number<W1>{}, Number<W2>{})
.Fold(I1, Number<K1>{}, Number<K2>{})
.Fold(I0, Number<N1>{}, Number<N2>{});
constexpr auto out_10d_thread_desc = fwd(out_k_h_w_n_thread_desc)
.Fold(I3, Number<1>{}, Number<N2>{})
.Fold(I2, Number<W1>{}, Number<1>{})
.Fold(I0, Number<1>{}, Number<K2>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"a: out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "a: out_10d_thread_desc");
print_ConstantTensorDescriptor(out_n_k_h_w_global_desc,
"a: out_n_k_h_w_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "a: out_10d_global_desc");
}
#endif
constexpr auto map_out_global2thread = Sequence<7, 8, 9, 0, 1, 2, 3, 4, 5, 6>{};
threadwise_tensor_slice_copy_reorder_given_dst2src_v2(
out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_n_k_h_w_global_desc.GetOffsetFromMultiIndex(
n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
out_10d_thread_desc.GetLengths(),
map_out_global2thread);
// Number<OutThreadCopyDataPerWrite_W>{});
}).Else([&](auto fwd) {
static_assert(fwd(GemmNPerThreadSubC) >= NPerBlock && NPerThread == NPerBlock &&
GemmNPerThreadSubC % NPerThread == 0,
"wrong!");
// output is a 10d tensor
constexpr index_t N1 = NPerBlock;
constexpr index_t W3 = GemmNPerThreadSubC / NPerBlock;
constexpr index_t W2 = GemmNLevel0Cluster * GemmNLevel1Cluster;
constexpr index_t W1 = WoPerBlock / fwd(W2 * W3);
constexpr index_t K2 = GemmMPerThreadSubC;
constexpr index_t K1 = KPerBlock / KPerThread;
constexpr auto out_10d_global_desc =
fwd(out_n_k_h_w_global_desc)
.Fold(I3, Number<W1>{}, Number<W2>{}, Number<W3>{})
.Fold(I1, Number<K1>{}, Number<K2>{})
.Fold(I0, Number<N1>{});
constexpr auto out_10d_thread_desc =
fwd(out_k_h_w_n_thread_desc)
.Fold(I3, Number<N1>{})
.Fold(I2, Number<W1>{}, Number<1>{}, Number<W3>{})
.Fold(I0, Number<1>{}, Number<K2>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
"b: out_k_h_w_n_thread_desc");
print_ConstantTensorDescriptor(out_10d_thread_desc, "b: out_10d_thread_desc");
print_ConstantTensorDescriptor(out_n_k_h_w_global_desc,
"b: out_n_k_h_w_global_desc");
print_ConstantTensorDescriptor(out_10d_global_desc, "b: out_10d_global_desc");
}
#endif
constexpr auto map_out_global2thread = Sequence<8, 9, 0, 1, 2, 3, 4, 5, 6, 7>{};
#if 0
threadwise_tensor_slice_copy_reorder_given_dst2src_v3(
out_10d_thread_desc,
p_out_thread,
out_10d_global_desc,
p_out_global +
out_n_k_h_w_global_desc.GetOffsetFromMultiIndex(
n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
out_10d_thread_desc.GetLengths(),
map_out_global2thread,
Number<OutThreadCopyDataPerWrite_W>{});
#else
threadwise_generic_tensor_slice_copy_v1(
out_10d_thread_desc.ReorderGivenNew2Old(map_out_global2thread),
p_out_thread,
make_zero_array<index_t, 10>(),
out_10d_global_desc,
p_out_global +
out_n_k_h_w_global_desc.GetOffsetFromMultiIndex(
n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
make_zero_array<index_t, 10>(),
out_10d_thread_desc.GetLengths().ReorderGivenNew2Old(map_out_global2thread),
arithmetic_sequence_gen<0, 10, 1>::type{},
Number<1>{});
#endif
});
}
};
} // namespace
#endif
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v2_chwn_cyxk_khwn.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V2_CHWN_CYXK_KHWN
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V2_CHWN_CYXK_KHWN
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_gemm.hpp"
namespace ck {
// define B = flatten(N, Hi, Wi)
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
index_t BPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t BPerThread,
index_t KPerThread,
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,
index_t InBlockCopyThreadPerDim0,
index_t InBlockCopyThreadPerDim1,
index_t WeiBlockCopyThreadPerDim0,
index_t WeiBlockCopyThreadPerDim1,
index_t InBlockCopyDataPerRead,
index_t WeiBlockCopyDataPerRead,
index_t OutThreadCopyDataPerWrite>
struct GridwiseConvolutionImplicitGemm_v2_chwn_cyxk_khwn
{
__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 in_chwn_global_desc = InGlobalDesc{};
constexpr auto wei_cyxk_global_desc = WeiGlobalDesc{};
constexpr auto out_khwn_global_desc = OutGlobalDesc{};
constexpr index_t C = in_chwn_global_desc.GetLength(I0);
constexpr index_t Hi = in_chwn_global_desc.GetLength(I1);
constexpr index_t Wi = in_chwn_global_desc.GetLength(I2);
constexpr index_t N = in_chwn_global_desc.GetLength(I3);
constexpr index_t K = out_khwn_global_desc.GetLength(I0);
constexpr index_t Ho = out_khwn_global_desc.GetLength(I1);
constexpr index_t Wo = out_khwn_global_desc.GetLength(I2);
constexpr index_t Y = wei_cyxk_global_desc.GetLength(I1);
constexpr index_t X = wei_cyxk_global_desc.GetLength(I2);
constexpr index_t B = N * Hi * Wi;
constexpr index_t BGhostRead = (Y - 1) * Wi + (X - 1);
// divide block work by 2d: [K, B]
constexpr index_t KBlockWork = (K + KPerBlock - 1) / KPerBlock;
constexpr index_t BBlockWork = (B + BPerBlock - 1) / BPerBlock;
const index_t k_block_work_id = get_block_1d_id() / BBlockWork;
const index_t b_block_work_id = get_block_1d_id() - k_block_work_id * BBlockWork;
const index_t k_block_data_begin = k_block_work_id * KPerBlock;
const index_t b_block_data_begin = b_block_work_id * BPerBlock;
// flattend (2d) tensor view of gridwise input
constexpr auto in_cb_global_desc = make_ConstantTensorDescriptor(Sequence<C, B>{});
constexpr auto wei_ek_global_desc = make_ConstantTensorDescriptor(Sequence<C * Y * X, K>{});
// tensor view of blockwise input and weight
// be careful of alignment
constexpr auto in_cb_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, BPerBlock + BGhostRead>{}, Number<InBlockCopyDataPerRead>{});
constexpr auto wei_ek_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock * Y * X, KPerBlock>{}, Number<WeiBlockCopyDataPerRead>{});
constexpr auto wei_cyxk_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, Y, X, KPerBlock>{}, Number<WeiBlockCopyDataPerRead>{});
// tensor view of threadwise output in register
constexpr auto out_kb_thread_desc =
make_ConstantTensorDescriptor(Sequence<KPerThread, BPerThread>{});
// blockwise in copy
// formmat is [CPerBlock,BPerBlock + BGhostRead]
#if 0
const auto blockwise_in_copy =
Blockwise2dTensorCopy1<BlockSize,
Float,
decltype(in_cb_global_desc),
decltype(in_cb_block_desc),
decltype(in_cb_block_desc.GetLengths())>{};
#elif 0
const auto blockwise_in_copy =
Blockwise2dTensorCopy2<BlockSize,
Float,
decltype(in_cb_global_desc),
decltype(in_cb_block_desc),
decltype(in_cb_block_desc.GetLengths()),
InBlockCopyThreadPerDim0,
InBlockCopyThreadPerDim1>{};
#elif 1
const auto blockwise_in_copy =
Blockwise2dTensorCopy3<BlockSize,
Float,
decltype(in_cb_global_desc),
decltype(in_cb_block_desc),
decltype(in_cb_block_desc.GetLengths()),
InBlockCopyDataPerRead>{};
#endif
// blockwise wei copy
// format is [CPerBlock*Y*X,KPerBlock]
#if 0
const auto blockwise_wei_copy =
Blockwise2dTensorCopy1<BlockSize,
Float,
decltype(wei_ek_global_desc),
decltype(wei_ek_block_desc),
decltype(wei_ek_block_desc.GetLengths())>{};
#elif 0
const auto blockwise_wei_copy =
Blockwise2dTensorCopy2<BlockSize,
Float,
decltype(wei_ek_global_desc),
decltype(wei_ek_block_desc),
decltype(wei_ek_block_desc.GetLengths()),
WeiBlockCopyThreadPerDim0,
WeiBlockCopyThreadPerDim1>{};
#elif 1
const auto blockwise_wei_copy =
Blockwise2dTensorCopy3<BlockSize,
Float,
decltype(wei_ek_global_desc),
decltype(wei_ek_block_desc),
decltype(wei_ek_block_desc.GetLengths()),
WeiBlockCopyDataPerRead>{};
#endif
// a series of blockwise GEMM
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx and b_mtx saved in LDS, c_mtx saved in register
// a_mtx[C,K] is a sub-matrix of wei_block[C,Y,X,K]
// b_mtx[C,B] is a subset of in_block[C,B + BGhostRead]
// c_mtx[K,B] is out_block[K,B]
constexpr auto a_cxk_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_cyxk_block_desc.GetStride(I0)>{});
constexpr auto b_cxb_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<BPerBlock>{}, Number<in_cb_block_desc.GetStride(I0)>{});
constexpr auto c_kxb_thread_mtx_desc =
make_ConstantMatrixDescriptor(Number<KPerThread>{}, Number<BPerThread>{});
const auto blockwise_gemm =
BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2<BlockSize,
decltype(a_cxk_block_mtx_desc),
decltype(b_cxb_block_mtx_desc),
decltype(c_kxb_thread_mtx_desc),
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB>{};
// LDS: be careful of alignment
constexpr index_t max_align =
math::lcm(index_t(4), InBlockCopyDataPerRead, WeiBlockCopyDataPerRead);
constexpr index_t in_block_space = in_cb_block_desc.GetElementSpace(Number<max_align>{});
constexpr index_t wei_block_space =
wei_cyxk_block_desc.GetElementSpace(Number<max_align>{});
__shared__ Float p_in_block[in_block_space];
__shared__ Float p_wei_block[wei_block_space];
const Float* p_in_global_block_offset =
p_in_global + in_cb_global_desc.GetOffsetFromMultiIndex(0, b_block_data_begin);
const Float* p_wei_global_block_offset =
p_wei_global +
wei_cyxk_global_desc.GetOffsetFromMultiIndex(0, 0, 0, k_block_data_begin);
// register
Float p_out_thread[out_kb_thread_desc.GetElementSpace()];
// set threadwise output to 0
threadwise_matrix_set_zero(c_kxb_thread_mtx_desc, p_out_thread);
for(index_t c_block_data_begin = 0; c_block_data_begin < C; c_block_data_begin += CPerBlock,
p_in_global_block_offset += CPerBlock * in_cb_global_desc.GetStride(I0),
p_wei_global_block_offset += CPerBlock * wei_cyxk_global_desc.GetStride(I0),
__syncthreads())
{
// load data
Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global_block_offset,
p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
p_wei_register_clipboard);
blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard, p_in_block);
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard, p_wei_block);
__syncthreads();
// compute on current data
// a series of GEMM
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
#if 1
blockwise_gemm.Run
#elif 0
blockwise_gemm.Run_RegisterDoubleBuffer
#elif 1
blockwise_gemm.Run_amd_asm
#endif
(p_wei_block + wei_cyxk_block_desc.GetOffsetFromMultiIndex(0, y, x, 0),
p_in_block + y * Wi + x,
p_out_thread);
}
}
}
// output: register to global mem,
const auto c_thread_mtx_begin =
blockwise_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t k_thread_data_begin = k_block_data_begin + c_thread_mtx_begin.row;
const index_t b_thread_data_begin = b_block_data_begin + c_thread_mtx_begin.col;
for(index_t k = 0; k < out_kb_thread_desc.GetLength(I0); ++k)
{
for(index_t b = 0; b < out_kb_thread_desc.GetLength(I1); ++b)
{
const auto c_thread_mtx_distance =
blockwise_gemm.GetDistanceFromBeginOfThreadMatrixC(k, b);
index_t k_data = k_thread_data_begin + c_thread_mtx_distance.row;
index_t b_data = b_thread_data_begin + c_thread_mtx_distance.col;
index_t h_data = b_data / (Wi * N);
index_t itmp = b_data - h_data * (Wi * N);
index_t w_data = itmp / N;
index_t n_data = itmp - w_data * N;
if(n_data < N && h_data < Ho && w_data < Wo)
{
p_out_global[out_khwn_global_desc.GetOffsetFromMultiIndex(
k_data, h_data, w_data, n_data)] =
p_out_thread[out_kb_thread_desc.GetOffsetFromMultiIndex(k, b)];
}
}
}
}
};
} // namespace ck
#endif
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v2_chwn_cyxk_khwn_lds_double_buffer.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V2_CHWN_CYXK_KHWN_LDS_DOUBLE_BUFFER
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V2_CHWN_CYXK_KHWN_LDS_DOUBLE_BUFFER
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "threadwise_tensor_slice_copy.hpp"
#include "blockwise_gemm.hpp"
namespace ck {
// define B = flatten(N, Hi, Wi)
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
index_t BPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t BPerThread,
index_t KPerThread,
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,
index_t InBlockCopyThreadPerDim0,
index_t InBlockCopyThreadPerDim1,
index_t WeiBlockCopyThreadPerDim0,
index_t WeiBlockCopyThreadPerDim1,
index_t InBlockCopyDataPerRead,
index_t WeiBlockCopyDataPerRead,
index_t OutThreadCopyDataPerWrite>
struct GridwiseConvolutionImplicitGemm_v2_chwn_cyxk_khwn_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
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_chwn_global_desc = InGlobalDesc{};
constexpr auto wei_cyxk_global_desc = WeiGlobalDesc{};
constexpr auto out_khwn_global_desc = OutGlobalDesc{};
constexpr index_t C = in_chwn_global_desc.GetLength(I0);
constexpr index_t Hi = in_chwn_global_desc.GetLength(I1);
constexpr index_t Wi = in_chwn_global_desc.GetLength(I2);
constexpr index_t N = in_chwn_global_desc.GetLength(I3);
constexpr index_t K = out_khwn_global_desc.GetLength(I0);
constexpr index_t Ho = out_khwn_global_desc.GetLength(I1);
constexpr index_t Wo = out_khwn_global_desc.GetLength(I2);
constexpr index_t Y = wei_cyxk_global_desc.GetLength(I1);
constexpr index_t X = wei_cyxk_global_desc.GetLength(I2);
constexpr index_t B = N * Hi * Wi;
constexpr index_t BGhostRead = (Y - 1) * Wi + (X - 1);
// assert for LDS double buffer
static_assert(C % (2 * CPerBlock) == 0, "C cannot be evenly divided");
// divide block work by 2d: [K, B]
constexpr index_t KBlockWork = (K + KPerBlock - 1) / KPerBlock;
constexpr index_t BBlockWork = (B + BPerBlock - 1) / BPerBlock;
const index_t k_block_work_id = get_block_1d_id() / BBlockWork;
const index_t b_block_work_id = get_block_1d_id() - k_block_work_id * BBlockWork;
const index_t k_block_data_begin = k_block_work_id * KPerBlock;
const index_t b_block_data_begin = b_block_work_id * BPerBlock;
// flattend (2d) tensor view of gridwise input
constexpr auto in_cb_global_desc = make_ConstantTensorDescriptor(Sequence<C, B>{});
constexpr auto wei_ek_global_desc = make_ConstantTensorDescriptor(Sequence<C * Y * X, K>{});
// tensor view of blockwise input and weight
// be careful of alignment
constexpr auto in_cb_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, BPerBlock + BGhostRead>{}, Number<InBlockCopyDataPerRead>{});
constexpr auto wei_ek_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock * Y * X, KPerBlock>{}, Number<WeiBlockCopyDataPerRead>{});
constexpr auto wei_cyxk_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, Y, X, KPerBlock>{}, Number<WeiBlockCopyDataPerRead>{});
// tensor view of threadwise output in register
constexpr auto out_kb_thread_desc =
make_ConstantTensorDescriptor(Sequence<KPerThread, BPerThread>{});
// blockwise in copy
// formmat is [CPerBlock,BPerBlock + BGhostRead]
#if 0
const auto blockwise_in_copy =
Blockwise2dTensorCopy1<BlockSize,
Float,
decltype(in_cb_global_desc),
decltype(in_cb_block_desc),
decltype(in_cb_block_desc.GetLengths())>{};
#elif 0
const auto blockwise_in_copy =
Blockwise2dTensorCopy2<BlockSize,
Float,
decltype(in_cb_global_desc),
decltype(in_cb_block_desc),
decltype(in_cb_block_desc.GetLengths()),
InBlockCopyThreadPerDim0,
InBlockCopyThreadPerDim1>{};
#elif 1
const auto blockwise_in_copy =
Blockwise2dTensorCopy3<BlockSize,
Float,
decltype(in_cb_global_desc),
decltype(in_cb_block_desc),
decltype(in_cb_block_desc.GetLengths()),
InBlockCopyDataPerRead>{};
#endif
// blockwise wei copy
// format is [CPerBlock*Y*X,KPerBlock]
#if 0
const auto blockwise_wei_copy =
Blockwise2dTensorCopy1<BlockSize,
Float,
decltype(wei_ek_global_desc),
decltype(wei_ek_block_desc),
decltype(wei_ek_block_desc.GetLengths())>{};
#elif 0
const auto blockwise_wei_copy =
Blockwise2dTensorCopy2<BlockSize,
Float,
decltype(wei_ek_global_desc),
decltype(wei_ek_block_desc),
decltype(wei_ek_block_desc.GetLengths()),
WeiBlockCopyThreadPerDim0,
WeiBlockCopyThreadPerDim1>{};
#elif 1
const auto blockwise_wei_copy =
Blockwise2dTensorCopy3<BlockSize,
Float,
decltype(wei_ek_global_desc),
decltype(wei_ek_block_desc),
decltype(wei_ek_block_desc.GetLengths()),
WeiBlockCopyDataPerRead>{};
#endif
// a series of blockwise GEMM
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx and b_mtx saved in LDS, c_mtx saved in register
// a_mtx[C,K] is a sub-matrix of wei_block[C,Y,X,K]
// b_mtx[C,B] is a subset of in_block[C,B + BGhostRead]
// c_mtx[K,B] is out_block[K,B]
constexpr auto a_cxk_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_cyxk_block_desc.GetStride(I0)>{});
constexpr auto b_cxb_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<BPerBlock>{}, Number<in_cb_block_desc.GetStride(I0)>{});
constexpr auto c_kxb_thread_mtx_desc =
make_ConstantMatrixDescriptor(Number<KPerThread>{}, Number<BPerThread>{});
const auto blockwise_gemm =
BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2<BlockSize,
decltype(a_cxk_block_mtx_desc),
decltype(b_cxb_block_mtx_desc),
decltype(c_kxb_thread_mtx_desc),
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB>{};
// LDS: be careful of alignment
constexpr index_t max_align =
math::lcm(index_t(4), InBlockCopyDataPerRead, WeiBlockCopyDataPerRead);
constexpr index_t in_block_space = in_cb_block_desc.GetElementSpace(Number<max_align>{});
constexpr index_t wei_block_space =
wei_cyxk_block_desc.GetElementSpace(Number<max_align>{});
// LDS double buffer
__shared__ Float p_in_block_double[2 * in_block_space];
__shared__ Float p_wei_block_double[2 * wei_block_space];
const Float* p_in_global_block_offset =
p_in_global + in_cb_global_desc.GetOffsetFromMultiIndex(0, b_block_data_begin);
const Float* p_wei_global_block_offset =
p_wei_global +
wei_cyxk_global_desc.GetOffsetFromMultiIndex(0, 0, 0, k_block_data_begin);
// preload data into LDS
{
Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global_block_offset,
p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
p_wei_register_clipboard);
blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard, p_in_block_double);
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard,
p_wei_block_double);
}
// register
Float p_out_thread[out_kb_thread_desc.GetElementSpace()];
// set threadwise output to 0
threadwise_matrix_set_zero(c_kxb_thread_mtx_desc, p_out_thread);
for(index_t c_block_data_begin = 0; c_block_data_begin + 2 * CPerBlock < C;
c_block_data_begin += 2 * CPerBlock)
{
#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;
// load next data
Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
p_in_global_block_offset += CPerBlock * in_cb_global_desc.GetStride(I0);
p_wei_global_block_offset += CPerBlock * wei_cyxk_global_desc.GetStride(I0);
__syncthreads();
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global_block_offset,
p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
p_wei_register_clipboard);
// compute on current data
// a series of GEMM
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
#if 1
blockwise_gemm.Run
#elif 0
blockwise_gemm.Run_RegisterDoubleBuffer
#elif 0
blockwise_gemm.Run_amd_asm
#endif
(p_wei_block_now +
wei_cyxk_block_desc.GetOffsetFromMultiIndex(0, y, x, 0),
p_in_block_now + y * Wi + x,
p_out_thread);
}
}
blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard,
p_in_block_next);
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard,
p_wei_block_next);
}
}
// tail
{
// even
p_in_global_block_offset += CPerBlock * in_cb_global_desc.GetStride(I0);
p_wei_global_block_offset += CPerBlock * wei_cyxk_global_desc.GetStride(I0);
__syncthreads();
Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
blockwise_in_copy.RunLoadRegisterClipboard(p_in_global_block_offset,
p_in_register_clipboard);
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
p_wei_register_clipboard);
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
#if 1
blockwise_gemm.Run
#elif 0
blockwise_gemm.Run_RegisterDoubleBuffer
#elif 0
blockwise_gemm.Run_amd_asm
#endif
(p_wei_block_double +
wei_cyxk_block_desc.GetOffsetFromMultiIndex(0, y, x, 0),
p_in_block_double + y * Wi + x,
p_out_thread);
}
}
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
__syncthreads();
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
#if 1
blockwise_gemm.Run
#elif 0
blockwise_gemm.Run_RegisterDoubleBuffer
#elif 0
blockwise_gemm.Run_amd_asm
#endif
(p_wei_block_double + wei_block_space +
wei_cyxk_block_desc.GetOffsetFromMultiIndex(0, y, x, 0),
p_in_block_double + in_block_space + y * Wi + x,
p_out_thread);
}
}
}
// output: register to global mem,
const auto c_thread_mtx_begin =
blockwise_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t k_thread_data_begin = k_block_data_begin + c_thread_mtx_begin.row;
const index_t b_thread_data_begin = b_block_data_begin + c_thread_mtx_begin.col;
if(Y == 1 && X == 1)
{ // pure 1x1 conv (non padding, 1x1 stride)
constexpr index_t K2_ = GemmMPerThreadSubC;
constexpr index_t K1_ = KPerBlock / KPerThread;
constexpr index_t B2_ = GemmNPerThreadSubC;
constexpr index_t B1_ = BPerBlock / BPerThread;
constexpr auto out_6d_global_desc = make_ConstantTensorDescriptor(
Sequence<K / (K1_ * K2_), K1_, K2_, B / (B1_ * B2_), B1_, B2_>{});
constexpr auto out_6d_thread_desc = make_ConstantTensorDescriptor(
Sequence<KPerBlock / (K1_ * K2_), 1, K2_, BPerBlock / (B1_ * B2_), 1, B2_>{});
constexpr auto out_kb_global_desc = make_ConstantTensorDescriptor(Sequence<K, B>{});
threadwise_6d_tensor_copy(out_6d_thread_desc,
p_out_thread,
out_6d_global_desc,
p_out_global +
out_kb_global_desc.GetOffsetFromMultiIndex(
k_thread_data_begin, b_thread_data_begin),
out_6d_thread_desc.GetLengths(),
Number<OutThreadCopyDataPerWrite>{});
}
else
{
for(index_t k = 0; k < out_kb_thread_desc.GetLength(I0); ++k)
{
for(index_t b = 0; b < out_kb_thread_desc.GetLength(I1); ++b)
{
const auto c_thread_mtx_distance =
blockwise_gemm.GetDistanceFromBeginOfThreadMatrixC(k, b);
index_t k_data = k_thread_data_begin + c_thread_mtx_distance.row;
index_t b_data = b_thread_data_begin + c_thread_mtx_distance.col;
index_t h_data = b_data / (Wi * N);
index_t itmp = b_data - h_data * (Wi * N);
index_t w_data = itmp / N;
index_t n_data = itmp - w_data * N;
if(n_data < N && h_data < Ho && w_data < Wo)
{
p_out_global[out_khwn_global_desc.GetOffsetFromMultiIndex(
k_data, h_data, w_data, n_data)] =
p_out_thread[out_kb_thread_desc.GetOffsetFromMultiIndex(k, b)];
}
}
}
}
}
};
} // namespace ck
#endif
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V3_NCHW_CYXK_NKHW
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V3_NCHW_CYXK_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"
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 CPerBlock,
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_C_N1_B_N2,
class InBlockCopyClusterLengths_C_N1_B_N2,
index_t InBlockCopySrcDataPerRead_B,
index_t InBlockCopyDstDataPerWrite_N2,
class WeiBlockCopySubLengths_C_K,
class WeiBlockCopyClusterLengths_C_K,
index_t WeiBlockCopyDataPerAccess_K>
struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_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 False = integral_constant<bool, false>{};
constexpr auto in_n_c_h_w_global_desc = InGlobalDesc{};
constexpr auto wei_c_y_x_k_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_c_y_x_k_global_desc.GetLength(I1);
constexpr index_t X = wei_c_y_x_k_global_desc.GetLength(I2);
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;
// divide block work by [K, B]
static_assert(K % KPerBlock == 0 && B % BPerBlock == 0 && C % CPerBlock == 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
// memory layout descriptor in device memory [N0, N1, N2, C, H, W]
constexpr auto in_n0_n1_n2_c_h_w_global_mem_desc =
in_n_c_h_w_global_desc.Fold(I0, Number<N1>{}, Number<N2>{});
// merged tensor descriptor in device memory [C, N1, B, N2], src of blockwise copy
constexpr auto in_c_n1_b_n2_global_merged_desc = make_ConstantMergedTensorDescriptor(
in_n0_n1_n2_c_h_w_global_mem_desc.Slice(I4, Number<Ho>{}).Slice(I5, Number<Wo>{}),
Sequence<3>{},
Sequence<1>{},
Sequence<0, 4, 5>{},
Sequence<2>{});
// memory layout descriptor in LDS [C, N1, B, N2], dst of blockwise copy
// be careful of LDS alignment
constexpr auto in_c_n1_b_n2_block_mem_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, N1, BPerBlock, N2>{}, Number<InBlockCopyDstDataPerWrite_N2>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with alignment
static_assert(in_c_n1_b_n2_block_mem_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_c_n1_b_n2_global_merged_desc),
decltype(in_c_n1_b_n2_block_mem_desc),
decltype(in_c_n1_b_n2_block_mem_desc.GetLengths()),
InBlockCopySubLengths_C_N1_B_N2,
InBlockCopyClusterLengths_C_N1_B_N2,
Sequence<0, 1, 3, 2>, // thread_arrange_order [C, N1, N2, B]
Sequence<1, 3, 0, 2>, // src_access_order [N1, N2, C, B]
Sequence<0, 1, 2, 3>, // dst_access_order [C, N1, B, N2]
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_c_k_global_desc = wei_c_y_x_k_global_desc.Extract(I0, I3);
// tensor descriptor in LDS, dst of blockwise copy
// be careful of LDS alignment
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDataPerAccess_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_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_C_K,
WeiBlockCopyClusterLengths_C_K,
Sequence<0, 1>, // thread_arrange_order [C, K]
Sequence<0, 1>, // src_access_order [C, K]
Sequence<0, 1>, // dst_access_order [C, K]
WeiBlockCopyDataPerAccess_K,
WeiBlockCopyDataPerAccess_K>(
{0, k_block_data_on_global}, {0, 0});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[CPerBlock, KPerBlock] is in LDS
// b_mtx[CPerBlocl, N1 * BPerBlock * N2] is in LDS
// c_mtx[KPerBlock, N1 * BPerBlock * N2] is distributed among threads, and saved in
// register
constexpr auto a_c_k_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_c_k_block_desc.GetStride(I0)>{});
constexpr auto b_c_n1bn2_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
Number<N1 * BPerBlock * N2>{},
Number<in_c_n1_b_n2_block_mem_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_c_k_block_mtx_desc),
decltype(b_c_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,
WeiBlockCopyDataPerAccess_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr index_t in_block_space =
in_c_n1_b_n2_block_mem_desc.GetElementSpace(Number<max_align>{});
constexpr index_t wei_block_space = wei_c_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);
#if 0
// do work
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
// calculate origin of block input and weight tensor on global memory
const Float* p_in_block_on_global =
p_in_global + in_n_c_h_w_global_desc.GetOffsetFromMultiIndex(0, 0, y, x);
const Float* p_wei_block_on_global =
p_wei_global + wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, x, 0);
for(index_t
c_block_data_on_global = 0;
c_block_data_on_global < C;
c_block_data_on_global += CPerBlock,
p_in_block_on_global += CPerBlock * in_n_c_h_w_global_desc.GetStride(I1),
p_wei_block_on_global += CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0))
{
blockwise_in_copy.Run(p_in_block_on_global, p_in_block);
blockwise_wei_copy.Run(p_wei_block_on_global, p_wei_block);
__syncthreads();
run_blockwise_gemm(p_wei_block, p_in_block, p_out_thread);
__syncthreads();
}
}
}
#else
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
// calculate origin of block input and weight tensor on global memory
const Float* p_in_block_on_global =
p_in_global + in_n_c_h_w_global_desc.GetOffsetFromMultiIndex(0, 0, y, x);
const Float* p_wei_block_on_global =
p_wei_global + wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, x, 0);
for(index_t c_block_data_on_global = 0; c_block_data_on_global < C;
c_block_data_on_global += CPerBlock)
{
blockwise_in_copy.Run(p_in_block_on_global, p_in_block);
blockwise_wei_copy.Run(p_wei_block_on_global, p_wei_block);
__syncthreads();
blockwise_gemm.Run(p_wei_block, p_in_block, p_out_thread);
__syncthreads();
blockwise_in_copy.MoveSlicingWindowOnSourceTensor(
I0, Number<CPerBlock>{}, True);
blockwise_wei_copy.MoveSlicingWindowOnSourceTensor(
I0, Number<CPerBlock>{}, True);
}
// reset C
blockwise_in_copy.MoveSlicingWindowOnSourceTensor(I0, Number<C>{}, False);
blockwise_wei_copy.MoveSlicingWindowOnSourceTensor(I0, Number<C>{}, False);
}
}
#endif
// 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
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v3_nchw_cyxk_nkhw_lds_double_buffer.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V3_NCHW_CYXK_NKHW_LDS_DOUBLE_BUFFER
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V3_NCHW_CYXK_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"
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 CPerBlock,
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_C_N1_B_N2,
class InBlockCopyClusterLengths_C_N1_B_N2,
index_t InBlockCopySrcDataPerRead_B,
index_t InBlockCopyDstDataPerWrite_N2,
class WeiBlockCopySubLengths_C_K,
class WeiBlockCopyClusterLengths_C_K,
index_t WeiBlockCopyDataPerAccess_K>
struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_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 I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
constexpr auto I6 = Number<6>{};
constexpr auto I7 = Number<7>{};
constexpr auto in_n_c_h_w_global_desc = InGlobalDesc{};
constexpr auto wei_c_y_x_k_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_c_y_x_k_global_desc.GetLength(I1);
constexpr index_t X = wei_c_y_x_k_global_desc.GetLength(I2);
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;
// divide block work by [K, B]
static_assert(K % KPerBlock == 0 && B % BPerBlock == 0 && C % (2 * CPerBlock) == 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
// memory layout descriptor in device memory [N0, N1, N2, C, H, W]
constexpr auto in_n0_n1_n2_c_h_w_global_mem_desc =
in_n_c_h_w_global_desc.Fold(I0, Number<N1>{}, Number<N2>{});
// merged tensor descriptor in device memory [C, N1, B, N2], src of blockwise copy
constexpr auto in_c_n1_b_n2_global_merged_desc = make_ConstantMergedTensorDescriptor(
in_n0_n1_n2_c_h_w_global_mem_desc.Slice(I4, Number<Ho>{}).Slice(I5, Number<Wo>{}),
Sequence<3>{},
Sequence<1>{},
Sequence<0, 4, 5>{},
Sequence<2>{});
// memory layout descriptor in LDS [C, N1, B, N2], dst of blockwise copy
// be careful of LDS alignment
constexpr auto in_c_n1_b_n2_block_mem_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, N1, BPerBlock, N2>{}, Number<InBlockCopyDstDataPerWrite_N2>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with alignment
static_assert(in_c_n1_b_n2_block_mem_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
const auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1<
BlockSize,
Float,
decltype(in_c_n1_b_n2_global_merged_desc),
decltype(in_c_n1_b_n2_block_mem_desc),
decltype(in_c_n1_b_n2_block_mem_desc.GetLengths()),
InBlockCopySubLengths_C_N1_B_N2,
InBlockCopyClusterLengths_C_N1_B_N2,
Sequence<0, 1, 3, 2>, // thread_arrange_order [C, N1, N2, B]
Sequence<1, 3, 0, 2>, // src_access_order [N1, N2, C, B]
Sequence<0, 1, 2, 3>, // dst_access_order [C, N1, B, N2]
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_c_k_global_desc = wei_c_y_x_k_global_desc.Extract(I0, I3);
// tensor descriptor in LDS, dst of blockwise copy
// be careful of LDS alignment
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<math::lcm(WeiBlockCopyDataPerAccess_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
const auto blockwise_wei_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
Float,
decltype(wei_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_C_K,
WeiBlockCopyClusterLengths_C_K,
Sequence<0, 1>, // thread_arrange_order [C, K]
Sequence<0, 1>, // src_access_order [C, K]
Sequence<0, 1>, // dst_access_order [C, K]
WeiBlockCopyDataPerAccess_K,
WeiBlockCopyDataPerAccess_K>(
{0, k_block_data_on_global}, {0, 0});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[CPerBlock, KPerBlock] is in LDS
// b_mtx[CPerBlocl, N1 * BPerBlock * N2] is in LDS
// c_mtx[KPerBlock, N1 * BPerBlock * N2] is distributed among threads, and saved in
// register
constexpr auto a_c_k_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_c_k_block_desc.GetStride(I0)>{});
constexpr auto b_c_n1bn2_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
Number<N1 * BPerBlock * N2>{},
Number<in_c_n1_b_n2_block_mem_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_c_k_block_mtx_desc),
decltype(b_c_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,
WeiBlockCopyDataPerAccess_K,
GemmDataPerReadA,
GemmDataPerReadB);
constexpr index_t in_block_space =
in_c_n1_b_n2_block_mem_desc.GetElementSpace(Number<max_align>{});
constexpr index_t wei_block_space = wei_c_k_block_desc.GetElementSpace(Number<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);
// do work
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
// calculate origin of block input and weight tensor on global memory
const Float* p_in_block_on_global =
p_in_global + in_n_c_h_w_global_desc.GetOffsetFromMultiIndex(0, 0, y, x);
const Float* p_wei_block_on_global =
p_wei_global + wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, x, 0);
// LDS double buffer: preload data into LDS
{
blockwise_in_copy.Run(p_in_block_on_global, p_in_block_double);
blockwise_wei_copy.Run(p_wei_block_on_global, p_wei_block_double);
}
// LDS double buffer: main body
for(index_t c_block_data_begin = 0; c_block_data_begin + 2 * CPerBlock < C;
c_block_data_begin += 2 * CPerBlock)
{
#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()];
p_in_block_on_global += CPerBlock * in_n_c_h_w_global_desc.GetStride(I1);
p_wei_block_on_global += CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy.RunLoadRegisterClipboard(p_in_block_on_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
run_blockwise_gemm(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
p_in_block_on_global += CPerBlock * in_n_c_h_w_global_desc.GetStride(I1);
p_wei_block_on_global += CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0);
__syncthreads();
// LDS doubel buffer: load next data from device mem
blockwise_in_copy.RunLoadRegisterClipboard(p_in_block_on_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
run_blockwise_gemm(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
run_blockwise_gemm(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;
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>{});
}
}
};
} // namesspace ck
#endif
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4_NCHW_KCYX_NKHW
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4_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_v4_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
composable_kernel/include/kernel_algorithm/gridwise_direct_convolution_2_vectorized_nchw_kcyx_nkhw.hpp deleted 100644 → 0
View file @ f0716f5b
#pragma once
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "blockwise_direct_convolution.hpp"
#include "threadwise_4d_tensor_op.hpp"
#include "threadwise_direct_convolution.hpp"
namespace ck {
template <class TInWei,
class TOut,
class TAccum,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
index_t ScalarPerVector,
index_t NPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
index_t NPerThread,
index_t KPerThread,
index_t CPerThread,
index_t HoPerThread,
index_t WoPerThread,
index_t InBlockCopyDataPerRead,
index_t WeiBlockCopyDataPerRead,
index_t BlockSize,
index_t GridSize>
__global__ void gridwise_direct_convolution_2_vectorized_nchw_kcyx_nkhw(
const typename vector_type<TInWei,
ScalarPerVector>::MemoryType* const __restrict__ p_in_vec_global,
const typename vector_type<TInWei,
ScalarPerVector>::MemoryType* const __restrict__ p_wei_vec_global,
TOut* const __restrict__ p_out_global)
{
using in_scalar_t = TInWei;
using in_vector_mem_t = typename vector_type<in_scalar_t, ScalarPerVector>::MemoryType;
using out_scalar_t = TOut;
using accum_t = TAccum;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_nchw_vec_global_desc = InGlobalDesc{};
constexpr auto wei_kcyx_vec_global_desc = WeiGlobalDesc{};
constexpr auto out_nkhw_global_desc = OutGlobalDesc{};
constexpr index_t N = in_nchw_vec_global_desc.GetLength(I0);
constexpr index_t K = wei_kcyx_vec_global_desc.GetLength(I0);
constexpr index_t C = wei_kcyx_vec_global_desc.GetLength(I1);
constexpr index_t Y = wei_kcyx_vec_global_desc.GetLength(I2);
constexpr index_t X = wei_kcyx_vec_global_desc.GetLength(I3);
constexpr auto wei_ke_vec_global_desc = make_ConstantTensorDescriptor(
Sequence<K, C * Y * X>{}); // 2d view of wei for blockwise copy
constexpr index_t HiPerBlock = HoPerBlock + Y - 1;
constexpr index_t WiPerBlock = WoPerBlock + X - 1;
constexpr auto in_nchw_vec_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<NPerBlock, CPerBlock, HiPerBlock, WiPerBlock>{}, Number<InBlockCopyDataPerRead>{});
constexpr auto wei_ke_vec_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<KPerBlock, CPerBlock * Y * X>{},
Number<WeiBlockCopyDataPerRead>{}); // 2d view of wei for blockwise copy
constexpr auto wei_kcyx_vec_block_desc =
make_ConstantTensorDescriptor(Sequence<KPerBlock, CPerBlock, Y, X>{},
Sequence<wei_ke_vec_block_desc.GetStride(I0), Y * X, X, 1>{});
// shared mem
constexpr index_t in_block_element_size =
in_nchw_vec_block_desc.GetElementSpace(Number<InBlockCopyDataPerRead>{});
constexpr index_t wei_block_element_size =
wei_kcyx_vec_block_desc.GetElementSpace(Number<WeiBlockCopyDataPerRead>{});
constexpr index_t max_align = InBlockCopyDataPerRead > WeiBlockCopyDataPerRead
? InBlockCopyDataPerRead
: WeiBlockCopyDataPerRead;
__shared__ in_vector_mem_t
p_in_vec_block[max_align * ((in_block_element_size + max_align - 1) / max_align)];
__shared__ in_vector_mem_t
p_wei_vec_block[max_align * ((wei_block_element_size + max_align - 1) / max_align)];
// threadwise tensors
constexpr index_t HiPerThread = HoPerThread + Y - 1;
constexpr index_t WiPerThread = WoPerThread + X - 1;
constexpr auto in_nchw_vec_thread_block_desc =
make_ConstantTensorDescriptor(Sequence<NPerThread, CPerThread, HiPerThread, WiPerThread>{},
in_nchw_vec_block_desc.GetStrides());
constexpr auto wei_kcyx_vec_thread_block_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread, CPerThread, Y, X>{}, wei_kcyx_vec_block_desc.GetStrides());
constexpr auto out_nkhw_thread_desc = get_convolution_output_default_4d_tensor_descriptor(
in_nchw_vec_thread_block_desc, wei_kcyx_vec_thread_block_desc);
// register
out_scalar_t p_out_thread[out_nkhw_thread_desc.GetElementSpace()];
// divide block work
constexpr index_t NBlockWork = (out_nkhw_global_desc.GetLength(I0) + NPerBlock - 1) / NPerBlock;
constexpr index_t KBlockWork = (out_nkhw_global_desc.GetLength(I1) + KPerBlock - 1) / KPerBlock;
constexpr index_t HBlockWork =
(out_nkhw_global_desc.GetLength(I2) + HoPerBlock - 1) / HoPerBlock;
constexpr index_t WBlockWork =
(out_nkhw_global_desc.GetLength(I3) + WoPerBlock - 1) / WoPerBlock;
const index_t block_id = blockIdx.x;
index_t itmp = block_id;
const index_t n_block_work_id = itmp / (KBlockWork * HBlockWork * WBlockWork);
itmp -= n_block_work_id * (KBlockWork * HBlockWork * WBlockWork);
const index_t k_block_work_id = itmp / (HBlockWork * WBlockWork);
itmp -= k_block_work_id * (HBlockWork * WBlockWork);
const index_t h_block_work_id = itmp / WBlockWork;
const index_t w_block_work_id = itmp - h_block_work_id * WBlockWork;
const index_t n_block_data_begin = n_block_work_id * NPerBlock;
const index_t k_block_data_begin = k_block_work_id * KPerBlock;
const index_t ho_block_data_begin = h_block_work_id * HoPerBlock;
const index_t wo_block_data_begin = w_block_work_id * WoPerBlock;
const index_t hi_block_data_begin = ho_block_data_begin; // minus padding
const index_t wi_block_data_begin = wo_block_data_begin; // minus padding
// divide thread work
constexpr index_t NThreadWork = (NPerBlock + NPerThread - 1) / NPerThread;
constexpr index_t KThreadWork = (KPerBlock + KPerThread - 1) / KPerThread;
constexpr index_t HThreadWork = (HoPerBlock + HoPerThread - 1) / HoPerThread;
constexpr index_t WThreadWork = (WoPerBlock + WoPerThread - 1) / WoPerThread;
const index_t thread_id = get_thread_local_1d_id();
itmp = thread_id;
const index_t n_thread_work_id = itmp / (KThreadWork * HThreadWork * WThreadWork);
itmp -= n_thread_work_id * (KThreadWork * HThreadWork * WThreadWork);
const index_t k_thread_work_id = itmp / (HThreadWork * WThreadWork);
itmp -= k_thread_work_id * (HThreadWork * WThreadWork);
const index_t h_thread_work_id = itmp / WThreadWork;
const index_t w_thread_work_id = itmp - h_thread_work_id * WThreadWork;
const index_t n_thread_data_begin = n_thread_work_id * NPerThread;
const index_t k_thread_data_begin = k_thread_work_id * KPerThread;
const index_t ho_thread_data_begin = h_thread_work_id * HoPerThread;
const index_t wo_thread_data_begin = w_thread_work_id * WoPerThread;
const index_t hi_thread_data_begin = ho_thread_data_begin;
const index_t wi_thread_data_begin = wo_thread_data_begin;
constexpr auto blockwise_in_copy =
Blockwise4dTensorCopy1<BlockSize,
in_vector_mem_t,
decltype(in_nchw_vec_global_desc),
decltype(in_nchw_vec_block_desc),
decltype(in_nchw_vec_block_desc.GetLengths()),
InBlockCopyDataPerRead>{};
#if 0
constexpr auto blockwise_wei_copy =
Blockwise4dTensorCopy1<BlockSize,
in_vector_mem_t,
decltype(wei_kcyx_vec_global_desc),
decltype(wei_kcyx_vec_block_desc),
decltype(wei_kcyx_vec_block_desc.GetLengths()),
1>{};
#elif 1
const auto blockwise_wei_copy =
Blockwise2dTensorCopy3<BlockSize,
in_vector_mem_t,
decltype(wei_ke_vec_global_desc),
decltype(wei_ke_vec_block_desc),
decltype(wei_ke_vec_block_desc.GetLengths()),
WeiBlockCopyDataPerRead>{};
#endif
#if 1 // debug
// set threadwise output tensor to 0
threadwise_4d_tensor_set_zero(out_nkhw_thread_desc, p_out_thread);
#endif
for(index_t c_block_data_begin = 0; c_block_data_begin < C;
c_block_data_begin += CPerBlock, __syncthreads())
{
// copy input tensor to LDS
blockwise_in_copy.Run(
p_in_vec_global +
in_nchw_vec_global_desc.GetOffsetFromMultiIndex(n_block_data_begin,
c_block_data_begin,
hi_block_data_begin,
wi_block_data_begin),
p_in_vec_block);
// copy weight tensor to LDS
blockwise_wei_copy.Run(p_wei_vec_global +
wei_kcyx_vec_global_desc.GetOffsetFromMultiIndex(
k_block_data_begin, c_block_data_begin, 0, 0),
p_wei_vec_block);
__syncthreads();
for(index_t c_thread_data = 0; c_thread_data < CPerBlock; c_thread_data += CPerThread)
{
// threadwise convolution
#if 1
threadwise_direct_convolution_2(
in_nchw_vec_thread_block_desc,
p_in_vec_block +
in_nchw_vec_block_desc.GetOffsetFromMultiIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
wei_kcyx_vec_thread_block_desc,
p_wei_vec_block +
wei_kcyx_vec_block_desc.GetOffsetFromMultiIndex(
k_thread_data_begin, c_thread_data, 0, 0),
out_nkhw_thread_desc,
p_out_thread);
#elif 0
threadwise_direct_convolution_3(
in_nchw_vec_thread_block_desc,
p_in_vec_block +
in_nchw_vec_block_desc.GetOffsetFromMultiIndex(n_thread_data_begin,
c_thread_data,
hi_thread_data_begin,
wi_thread_data_begin),
wei_kcyx_vec_thread_block_desc,
p_wei_vec_block +
wei_kcyx_vec_block_desc.GetOffsetFromMultiIndex(
k_thread_data_begin, c_thread_data, 0, 0),
out_nkhw_thread_desc,
p_out_thread);
#endif
}
}
// copy output tensor from register to global mem
threadwise_4d_tensor_copy(out_nkhw_thread_desc,
p_out_thread,
out_nkhw_global_desc,
p_out_global +
out_nkhw_global_desc.GetOffsetFromMultiIndex(
n_block_data_begin + n_thread_data_begin,
k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin),
out_nkhw_thread_desc.GetLengths());
}
} // namespace ck
composable_kernel/include/kernel_algorithm/gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded.hpp deleted 100644 → 0
View file @ f0716f5b
#pragma once
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "threadwise_4d_tensor_op.hpp"
#include "blockwise_gemm.hpp"
namespace ck {
template <index_t GridSize,
index_t BlockSize,
class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
class LowerPads,
class UpperPads,
index_t NPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
index_t NPerThread,
index_t KPerThread,
index_t CPerThread,
index_t HoPerThread,
index_t WoPerThread,
index_t WeiBlockCopyThreadPerDim0,
index_t WeiBlockCopyThreadPerDim1>
__global__ void gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded(
const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global)
{
// NPerThread == NPerBlock, because the format of input in LDS [C,Hi,Wi,N]
// for GEMM trans([C,K]) * [C,Wo*N], we need a thread to do all the "N"
// if we use [C,Hi,N,Wi,N] in LDS, then NPerThread can be different from NPerBlock
static_assert(NPerBlock % NPerThread == 0, "wrong! NPerBlock % NPerThread !=0");
static_assert((NPerThread < NPerBlock && WoPerThread == 1) || NPerThread == NPerBlock,
"wrong!");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_chwn_global_desc = InGlobalDesc{};
constexpr auto wei_cyxk_global_desc = WeiGlobalDesc{};
constexpr auto out_khwn_global_desc = OutGlobalDesc{};
constexpr index_t C = in_chwn_global_desc.GetLength(I0);
constexpr index_t K = out_khwn_global_desc.GetLength(I0);
constexpr index_t Ho = out_khwn_global_desc.GetLength(I1);
constexpr index_t Wo = out_khwn_global_desc.GetLength(I2);
constexpr index_t N = out_khwn_global_desc.GetLength(I3);
constexpr index_t Y = wei_cyxk_global_desc.GetLength(I1);
constexpr index_t X = wei_cyxk_global_desc.GetLength(I2);
constexpr index_t HPadLow = LowerPads{}.Get(I0);
constexpr index_t WPadLow = LowerPads{}.Get(I1);
constexpr index_t HPadUp = UpperPads{}.Get(I0);
constexpr index_t WPadUp = UpperPads{}.Get(I1);
constexpr index_t HiPerBlock = HoPerBlock + Y - 1;
constexpr index_t WiPerBlock = WoPerBlock + X - 1;
// divide block work: [K, Ho, Wo, N]
constexpr index_t KBlockWork = (K + KPerBlock - 1) / KPerBlock;
constexpr index_t HBlockWork = (Ho + HoPerBlock - 1) / HoPerBlock;
constexpr index_t WBlockWork = (Wo + WoPerBlock - 1) / WoPerBlock;
constexpr index_t NBlockWork = (N + NPerBlock - 1) / NPerBlock;
const index_t k_block_work_id = get_block_1d_id() / (HBlockWork * WBlockWork * NBlockWork);
index_t itmp = get_block_1d_id() - k_block_work_id * (HBlockWork * WBlockWork * NBlockWork);
const index_t h_block_work_id = itmp / (WBlockWork * NBlockWork);
itmp -= h_block_work_id * (WBlockWork * NBlockWork);
const index_t w_block_work_id = itmp / NBlockWork;
const index_t n_block_work_id = itmp - w_block_work_id * NBlockWork;
const index_t k_block_data_begin = k_block_work_id * KPerBlock;
const index_t ho_block_data_begin = h_block_work_id * HoPerBlock;
const index_t wo_block_data_begin = w_block_work_id * WoPerBlock;
const index_t n_block_data_begin = n_block_work_id * NPerBlock;
// flattened (2d) tensor view of wei in global mem
constexpr auto wei_ek_global_desc = make_ConstantTensorDescriptor(Sequence<C * Y * X, K>{});
// tensor view of blockwise input and weight in LDS
constexpr auto in_chwn_block_desc =
make_ConstantTensorDescriptor(Sequence<CPerBlock, HiPerBlock, WiPerBlock, NPerBlock>{});
constexpr auto wei_cyxk_block_desc =
make_ConstantTensorDescriptor(Sequence<CPerBlock, Y, X, KPerBlock>{});
// flattened (2d) tensor view of wei in LDS
constexpr auto wei_ek_block_desc =
make_ConstantTensorDescriptor(Sequence<CPerBlock * Y * X, KPerBlock>{});
// tensor view of threadwise output in register
constexpr auto out_hkwn_thread_desc =
make_ConstantTensorDescriptor(Sequence<HoPerThread, KPerThread, WoPerThread, NPerThread>{});
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_chwn_block_desc, "in_chwn_block_desc");
print_ConstantTensorDescriptor(wei_cyxk_block_desc, "wei_cyxk_block_desc");
print_ConstantTensorDescriptor(out_hkwn_thread_desc, "out_hkwn_thread_desc");
}
#endif
// blockwise copy
// input: format is [C, Hi, Wi, N]
const index_t h_block_pad_low = h_block_work_id == 0 ? HPadLow : 0;
const index_t w_block_pad_low = w_block_work_id == 0 ? WPadLow : 0;
const index_t h_block_pad_up = h_block_work_id == HBlockWork - 1 ? HPadUp : 0;
const index_t w_block_pad_up = w_block_work_id == WBlockWork - 1 ? WPadUp : 0;
#if 0
if(get_thread_local_1d_id() == 0)
;
{
printf(
"%u %u, h_block_pad_low %u w_block_pad_low %u h_block_pad_up %u w_block_pad_up %u\n",
get_block_1d_id(),
get_thread_local_1d_id(),
h_block_pad_low,
w_block_pad_low,
h_block_pad_up,
w_block_pad_up);
}
#endif
constexpr auto blockwise_in_copy =
BlockwiseChwnTensorCopyPadded<BlockSize,
Float,
decltype(in_chwn_global_desc),
decltype(in_chwn_block_desc),
decltype(in_chwn_block_desc.GetLengths()),
LowerPads>{};
#if 0
// weight: format is [C,Y,X,K]
constexpr auto blockwise_wei_copy =
Blockwise4dTensorCopy1<BlockSize,
Float,
decltype(wei_cyxk_global_desc),
decltype(wei_cyxk_block_desc),
decltype(wei_cyxk_block_desc.GetLengths())>{};
#elif 0
// weight: format is [C*Y*X,K]
constexpr auto blockwise_wei_copy =
Blockwise2dTensorCopy1<BlockSize,
Float,
decltype(wei_ek_global_desc),
decltype(wei_ek_block_desc),
decltype(wei_ek_block_desc.GetLengths())>{};
#elif 1
// weight: format is [C*Y*X,K]
const auto blockwise_wei_copy = Blockwise2dTensorCopy2<BlockSize,
Float,
decltype(wei_ek_global_desc),
decltype(wei_ek_block_desc),
decltype(wei_ek_block_desc.GetLengths()),
WeiBlockCopyThreadPerDim0,
WeiBlockCopyThreadPerDim1>{};
#endif
// a series of blockwise batched GEMM
// C_matrix += transpose(A_matrix) * B_matrix
// A_matrix and B_matrix saved in LDS, C_matrix saved in register
// A_matrix[C,K] is a sub-matrix of wei_block[C,Y,X,K]
// B_matrix[C,Wo*N] is a sub-matrix of in_block[C,Hi,Wi,N]
// C_matrix[K,Wo*N] is a sub-matrix of out_block[Ho,K,Wo,N]
constexpr auto a_cxk_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_cyxk_block_desc.GetStride(I0)>{});
constexpr auto b_cxwn_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
Number<WoPerBlock * NPerBlock>{},
Number<in_chwn_block_desc.GetStride(I0)>{});
constexpr auto c_kxwn_thread_mtx_desc =
make_ConstantMatrixDescriptor(Number<KPerThread>{}, Number<WoPerThread * NPerThread>{});
const auto blockwise_batch_gemm =
Blockwise1dStridedBatchedGemmBlockABlockBThreadC<BlockSize,
decltype(a_cxk_block_mtx_desc),
decltype(b_cxwn_block_mtx_desc),
decltype(c_kxwn_thread_mtx_desc),
true,
false,
false,
0,
in_chwn_block_desc.GetStride(I1),
out_hkwn_thread_desc.GetStride(I0),
HoPerBlock,
HoPerThread,
CPerThread,
true>{};
// LDS
constexpr index_t in_block_element_size = in_chwn_block_desc.GetElementSpace();
constexpr index_t wei_block_element_size = wei_cyxk_block_desc.GetElementSpace();
__shared__ Float p_in_block[in_block_element_size];
__shared__ Float p_wei_block[wei_block_element_size];
// register
Float p_out_thread[out_hkwn_thread_desc.GetElementSpace()];
// set threadwise output tensor to 0
threadwise_4d_tensor_set_zero(out_hkwn_thread_desc, p_out_thread);
const Float* p_wei_global_block_begin =
p_wei_global + wei_ek_global_desc.GetOffsetFromMultiIndex(0, k_block_data_begin);
for(index_t c_block_data_begin = 0; c_block_data_begin < C; c_block_data_begin += CPerBlock,
p_wei_global_block_begin += CPerBlock * wei_ek_global_desc.GetStride(I0),
__syncthreads())
{
#if 1
// input: global mem to LDS,
blockwise_in_copy.Run(p_in_global,
c_block_data_begin,
ho_block_data_begin,
wo_block_data_begin,
n_block_data_begin,
p_in_block,
h_block_pad_low,
w_block_pad_low,
h_block_pad_up,
w_block_pad_up);
#endif
#if 1
// weight: global mem to LDS,
blockwise_wei_copy.Run(p_wei_global_block_begin, p_wei_block);
#endif
__syncthreads();
// a series of batched GEMM
for(index_t y = 0; y < Y; ++y)
{
for(index_t x = 0; x < X; ++x)
{
auto f_accum = [](auto& acc, const auto&& v) { acc += v; };
blockwise_batch_gemm.Run(
p_wei_block + wei_cyxk_block_desc.GetOffsetFromMultiIndex(0, y, x, 0),
p_in_block + in_chwn_block_desc.GetOffsetFromMultiIndex(0, y, x, 0),
p_out_thread,
f_accum);
}
}
}
const auto matrix_c_index =
blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t ho_thread_data_begin = matrix_c_index.batch;
const index_t k_thread_data_begin = matrix_c_index.row;
const index_t wo_thread_data_begin = matrix_c_index.col / NPerBlock;
const index_t n_thread_data_begin = matrix_c_index.col - wo_thread_data_begin * NPerBlock;
#if 0
printf("block %u %u, %u %u %u %u, %u %u %u %u, %f \n",
get_block_1d_id(), get_thread_local_1d_id(),
ho_block_data_begin, k_block_data_begin, wo_block_data_begin, n_block_data_begin,
ho_thread_data_begin, k_thread_data_begin, wo_thread_data_begin, n_thread_data_begin,
p_out_thread[0]);
#endif
// output: register to global mem,
// convert out_thread[Ho,K,Wo,N] to out_global[K,Ho,Wo,N]
constexpr auto reorder_khwn_from_hkwn = Sequence<1, 0, 2, 3>{};
threadwise_4d_tensor_copy_reorder_by_get_dst_from_src(
out_hkwn_thread_desc,
p_out_thread,
out_khwn_global_desc,
p_out_global +
out_khwn_global_desc.GetOffsetFromMultiIndex(k_block_data_begin + k_thread_data_begin,
ho_block_data_begin + ho_thread_data_begin,
wo_block_data_begin + wo_thread_data_begin,
n_block_data_begin + n_thread_data_begin),
out_hkwn_thread_desc.GetLengths(),
reorder_khwn_from_hkwn);
}
} // namespace ck
composable_kernel/include/tensor_operation/blockwise_2d_tensor_op.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_BLOCKWISE_2D_TENSOR_OP_HPP
#define CK_BLOCKWISE_2D_TENSOR_OP_HPP
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
namespace ck {
template <index_t BlockSize, class Float, class DstDesc, class F>
__device__ void
blockwise_2d_tensor_pointwise_operation_unary(DstDesc, Float* __restrict__ p_dst, F f)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto dst_desc = DstDesc{};
constexpr auto desc = make_ConstantTensorDescriptor(dst_desc.GetLengths());
#if 0
if(get_thread_local_1d_id() == 0)
{
print_ConstantTensorDescriptor(dst_desc, "blockwise_4d_tensor_op_unary: dst_desc: ");
print_ConstantTensorDescriptor(desc, "blockwise_4d_tensor_op_unary: desc: ");
}
#endif
constexpr index_t NLoop = desc.GetElementSize() / BlockSize;
for(index_t iloop = 0; iloop < NLoop; ++iloop)
{
index_t is = get_thread_local_1d_id() + iloop * BlockSize;
const index_t did0 = is / desc.GetStride(I0);
is -= did0 * desc.GetStride(I0);
const index_t did1 = is / desc.GetStride(I1);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1);
f(p_dst[dindex]);
}
constexpr bool has_tail = (desc.GetElementSize() > NLoop * BlockSize);
if(has_tail)
{
index_t is = get_thread_local_1d_id() + NLoop * BlockSize;
if(is < desc.GetElementSize())
{
const index_t did0 = is / desc.GetStride(I0);
is -= did0 * desc.GetStride(I0);
const index_t did1 = is / desc.GetStride(I1);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1);
f(p_dst[dindex]);
}
}
}
// Function: p_dst[reorder[i0], reorder[i1] = p_src[i0,i1]
// TODO: in order to optimize mem access for different mem type,
// need to write specialized version
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class SrcOpLengths,
class MapDst2Src,
class F>
__device__ void blockwise_2d_tensor_pointwise_operation_binary_reorder_by_get_dst_from_src(
SrcDesc,
const Float* __restrict__ p_src,
DstDesc,
Float* __restrict__ p_dst,
SrcOpLengths,
MapDst2Src,
F f)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr index_t IR0 = MapDst2Src{}.Get(I0);
constexpr index_t IR1 = MapDst2Src{}.Get(I1);
constexpr auto src_desc = SrcDesc{};
constexpr auto dst_desc = DstDesc{};
constexpr auto ref_desc = make_ConstantTensorDescriptor(SrcOpLengths{});
constexpr index_t NLoop = ref_desc.GetElementSize() / BlockSize;
for(index_t iloop = 0; iloop < NLoop; ++iloop)
{
index_t is = get_thread_local_1d_id() + iloop * BlockSize;
index_t did[2];
did[0] = is / ref_desc.GetStride(I0);
is -= did[0] * ref_desc.GetStride(I0);
did[1] = is / ref_desc.GetStride(I1);
const index_t aindex = src_desc.GetOffsetFromMultiIndex(did[0], did[1]);
const index_t bindex = dst_desc.GetOffsetFromMultiIndex(did[IR0], did[IR1]);
f(p_src[aindex], p_dst[bindex]);
}
constexpr bool has_tail = (ref_desc.GetElementSize() > NLoop * BlockSize);
if(has_tail)
{
index_t is = get_thread_local_1d_id() + NLoop * BlockSize;
if(is < ref_desc.GetElementSize())
{
index_t did[2];
did[0] = is / ref_desc.GetStride(I0);
is -= did[0] * ref_desc.GetStride(I0);
did[1] = is / ref_desc.GetStride(I1);
const index_t aindex = src_desc.GetOffsetFromMultiIndex(did[0], did[1]);
const index_t bindex = dst_desc.GetOffsetFromMultiIndex(did[IR0], did[IR1]);
f(p_src[aindex], p_dst[bindex]);
}
}
}
template <index_t BlockSize, class Float, class DstDesc>
__device__ void blockwise_2d_tensor_set_zero(DstDesc, Float* __restrict__ p_dst)
{
auto f_set_zero = [](Float& v) { v = Float(0); };
blockwise_2d_tensor_pointwise_operation_unary<BlockSize>(DstDesc{}, p_dst, f_set_zero);
}
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class SrcOpLengths,
class MapDst2Src>
__device__ void
blockwise_2d_tensor_copy_reorder_by_get_dst_from_src(SrcDesc,
const Float* __restrict__ p_src,
DstDesc,
Float* __restrict__ p_dst,
SrcOpLengths,
MapDst2Src)
{
auto f_copy = [](const Float& src, Float& dst) { dst = src; };
blockwise_2d_tensor_pointwise_operation_binary_reorder_by_get_dst_from_src<BlockSize>(
SrcDesc{}, p_src, DstDesc{}, p_dst, SrcOpLengths{}, MapDst2Src{}, f_copy);
}
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class CopyLengths,
index_t DataPerRead>
struct Blockwise2dTensorCopy1
{
using vector_t = typename vector_type<Float, DataPerRead>::MemoryType;
__device__ constexpr Blockwise2dTensorCopy1()
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
static_assert(DataPerRead == 1 ||
(SrcDesc{}.GetStride(I1) == 1 && DstDesc{}.GetStride(I1) == 1),
"wrong! only support stride1 == 1 if DataPerRead > 1!\n");
static_assert(DataPerRead == 1 || DataPerRead == 2 || DataPerRead == 4,
"wrong! only support DataPerRead == 1, 2 or 4!\n");
static_assert(SrcDesc{}.GetStride(I0) % DataPerRead == 0 &&
DstDesc{}.GetStride(I0) % DataPerRead == 0,
"src and dst stride2 should be multiple of DataPerRead to keep alignment");
// we allow out-of-bound read from src in D1 dimension,
// but we need to make sure dst stride0 is big enough,
// so that the out-of-bound write won't contaminate next line in dst
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t read_per_d1 = math::integer_divide_ceil(L1, DataPerRead);
static_assert(read_per_d1 * DataPerRead <= DstDesc{}.GetStride(I0),
"wrong! out-of-bound write will contaminate next line!\n");
}
__device__ void Run(const Float* __restrict__ p_src, Float* __restrict__ p_dst) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto src_desc = SrcDesc{};
constexpr auto dst_desc = DstDesc{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t read_per_d1 = math::integer_divide_ceil(L1, DataPerRead);
constexpr auto ref_desc = make_ConstantTensorDescriptor(Sequence<L0, read_per_d1>{});
constexpr index_t NLoop = ref_desc.GetElementSize() / BlockSize;
auto f_copy = [&](index_t is) {
index_t did[4];
did[0] = is / ref_desc.GetStride(I0);
is -= did[0] * ref_desc.GetStride(I0);
did[1] = is / ref_desc.GetStride(I1);
const index_t src_index =
src_desc.GetOffsetFromMultiIndex(did[0], did[1] * DataPerRead);
const index_t dst_index =
dst_desc.GetOffsetFromMultiIndex(did[0], did[1] * DataPerRead);
*(reinterpret_cast<vector_t*>(p_dst + dst_index)) =
*(reinterpret_cast<const vector_t*>(p_src + src_index));
};
for(index_t iloop = 0; iloop < NLoop; ++iloop)
{
index_t is = get_thread_local_1d_id() + iloop * BlockSize;
f_copy(is);
}
constexpr bool has_tail = (ref_desc.GetElementSize() > NLoop * BlockSize);
if(has_tail)
{
index_t is = get_thread_local_1d_id() + NLoop * BlockSize;
if(is < ref_desc.GetElementSize())
{
f_copy(is);
}
}
}
};
// need to be aligned to float4 and float2
// stride1 need to be 1 for both source and destination
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class SrcOpLengths,
index_t ThreadPerDim0,
index_t ThreadPerDim1>
struct Blockwise2dTensorCopy2
{
index_t mThreadId0;
index_t mThreadId1;
__device__ Blockwise2dTensorCopy2()
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
static_assert(SrcDesc{}.GetStride(I1) == 1 && DstDesc{}.GetStride(I1) == 1,
"wrong! stride is not 1!\n");
mThreadId0 = get_thread_local_1d_id() / ThreadPerDim1;
mThreadId1 = get_thread_local_1d_id() - mThreadId0 * ThreadPerDim1;
}
__device__ void Run(const Float* __restrict__ p_src, Float* __restrict__ p_dst) const
{
static_assert(is_same<Float, float>{}, "wrong! only support float!\n");
using Float4 = float4;
using Float2 = float2;
if(get_thread_local_1d_id() >= ThreadPerDim0 * ThreadPerDim1)
return;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto src_desc = SrcDesc{};
constexpr auto dst_desc = DstDesc{};
// check alignment
constexpr bool align_v4 =
src_desc.GetStride(I0) % 4 == 0 && dst_desc.GetStride(I0) % 4 == 0;
constexpr bool align_v2 =
src_desc.GetStride(I0) % 2 == 0 && dst_desc.GetStride(I0) % 2 == 0;
constexpr index_t L0 = SrcOpLengths{}.Get(I0);
constexpr index_t L1 = SrcOpLengths{}.Get(I1);
constexpr index_t Dim0Loop = L0 / ThreadPerDim0;
constexpr bool d0_has_tail = (L0 > ThreadPerDim0 * Dim0Loop);
constexpr index_t Dim1V4Loop = align_v4 ? L1 / (ThreadPerDim1 * 4) : 0;
constexpr index_t Dim1V2Loop =
align_v2 ? (L1 - Dim1V4Loop * (ThreadPerDim1 * 4)) / (ThreadPerDim1 * 2) : 0;
constexpr index_t Dim1V1Loop =
(L1 - Dim1V4Loop * (ThreadPerDim1 * 4) - Dim1V2Loop * (ThreadPerDim1 * 2)) /
ThreadPerDim1;
constexpr bool d1_has_tail =
(L1 > ThreadPerDim1 * (4 * Dim1V4Loop + 2 * Dim1V2Loop + Dim1V1Loop));
for(index_t d0loop = 0; d0loop < Dim0Loop; ++d0loop)
{
index_t did0 = d0loop * ThreadPerDim0 + mThreadId0;
// v4
for(index_t d1v4loop = 0; d1v4loop < Dim1V4Loop; ++d1v4loop)
{
index_t did1 = d1v4loop * 4 * ThreadPerDim1 + 4 * mThreadId1;
const index_t sindex = src_desc.GetOffsetFromMultiIndex(did0, did1);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1);
*(reinterpret_cast<Float4*>(p_dst + dindex)) =
*(reinterpret_cast<const Float4*>(p_src + sindex));
}
// v2
for(index_t d1v2loop = 0; d1v2loop < Dim1V2Loop; ++d1v2loop)
{
index_t did1 =
Dim1V4Loop * 4 * ThreadPerDim1 + d1v2loop * 2 * ThreadPerDim1 + 2 * mThreadId1;
const index_t sindex = src_desc.GetOffsetFromMultiIndex(did0, did1);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1);
*(reinterpret_cast<Float2*>(p_dst + dindex)) =
*(reinterpret_cast<const Float2*>(p_src + sindex));
}
// v1
for(index_t d1v1loop = 0; d1v1loop < Dim1V1Loop; ++d1v1loop)
{
index_t did1 = Dim1V4Loop * 4 * ThreadPerDim1 + Dim1V2Loop * 2 * ThreadPerDim1 +
d1v1loop * ThreadPerDim1 + mThreadId1;
const index_t sindex = src_desc.GetOffsetFromMultiIndex(did0, did1);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1);
p_dst[dindex] = p_src[sindex];
}
// dim-1 tail
if(d1_has_tail)
{
index_t did1 = Dim1V4Loop * 4 * ThreadPerDim1 + Dim1V2Loop * 2 * ThreadPerDim1 +
Dim1V1Loop * ThreadPerDim1 + mThreadId1;
if(did1 < L1)
{
const index_t sindex = src_desc.GetOffsetFromMultiIndex(did0, did1);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1);
p_dst[dindex] = p_src[sindex];
}
}
}
// dim-0 tail
if(d0_has_tail)
{
index_t did0 = Dim0Loop * ThreadPerDim0 + mThreadId0;
if(did0 < L0)
{
// v4
for(index_t d1v4loop = 0; d1v4loop < Dim1V4Loop; ++d1v4loop)
{
index_t did1 = d1v4loop * 4 * ThreadPerDim1 + 4 * mThreadId1;
const index_t sindex = src_desc.GetOffsetFromMultiIndex(did0, did1);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1);
*(reinterpret_cast<Float4*>(p_dst + dindex)) =
*(reinterpret_cast<const Float4*>(p_src + sindex));
}
// v2
for(index_t d1v2loop = 0; d1v2loop < Dim1V2Loop; ++d1v2loop)
{
index_t did1 = Dim1V4Loop * 4 * ThreadPerDim1 + d1v2loop * 2 * ThreadPerDim1 +
2 * mThreadId1;
const index_t sindex = src_desc.GetOffsetFromMultiIndex(did0, did1);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1);
*(reinterpret_cast<Float2*>(p_dst + dindex)) =
*(reinterpret_cast<const Float2*>(p_src + sindex));
}
// v1
for(index_t d1v1loop = 0; d1v1loop < Dim1V1Loop; ++d1v1loop)
{
index_t did1 = Dim1V4Loop * 4 * ThreadPerDim1 + Dim1V2Loop * 2 * ThreadPerDim1 +
d1v1loop * ThreadPerDim1 + mThreadId1;
const index_t sindex = src_desc.GetOffsetFromMultiIndex(did0, did1);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1);
p_dst[dindex] = p_src[sindex];
}
// tail
if(d1_has_tail)
{
index_t did1 = Dim1V4Loop * 4 * ThreadPerDim1 + Dim1V2Loop * 2 * ThreadPerDim1 +
Dim1V1Loop * ThreadPerDim1 + mThreadId1;
if(did1 < L1)
{
const index_t sindex = src_desc.GetOffsetFromMultiIndex(did0, did1);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1);
p_dst[dindex] = p_src[sindex];
}
}
}
}
}
};
// starting point need to be aligned to float4 or float2 or float
// stride1 need to be 1 for both source and destination
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class CopyLengths,
index_t DataPerRead>
struct Blockwise2dTensorCopy3
{
using vector_t = typename vector_type<Float, DataPerRead>::MemoryType;
index_t mSrcMyThreadOffset;
index_t mDstMyThreadOffset;
__device__ Blockwise2dTensorCopy3(Array<index_t, 2> src_block_data_multi_id_begin,
Array<index_t, 2> dst_block_data_multi_id_begin)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
static_assert(DataPerRead == 1 ||
(SrcDesc{}.GetStride(I1) == 1 && DstDesc{}.GetStride(I1) == 1),
"wrong! only support stride1 == 1 if DataPerRead > 1!\n");
static_assert(DataPerRead == 1 || DataPerRead == 2 || DataPerRead == 4,
"wrong! only support DataPerRead == 1, 2 or 4!\n");
static_assert(SrcDesc{}.GetStride(I0) % DataPerRead == 0 &&
DstDesc{}.GetStride(I0) % DataPerRead == 0,
"src and dst stride should be multiple of DataPerRead to keep alignment");
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t thread_per_d1 = (L1 + DataPerRead - 1) / DataPerRead;
constexpr index_t thread_per_d0 = BlockSize / thread_per_d1;
// we allow out-of-bound read from src in D1 dimension,
// but we need to make sure dst stride is big enough,
// so that the out-of-bound write won't contaminate next line in dst
static_assert(thread_per_d1 * DataPerRead <= DstDesc{}.GetStride(I0),
"wrong! out-of-bound write will contaminate next line!\n");
static_assert(thread_per_d0 >= 1, "wrong! not enough threads to cover one line\n");
constexpr index_t num_active_thread = thread_per_d0 * thread_per_d1;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
const index_t thread_id_d0 = get_thread_local_1d_id() / thread_per_d1;
const index_t thread_id_d1 = get_thread_local_1d_id() - thread_id_d0 * thread_per_d1;
mSrcMyThreadOffset = SrcDesc{}.GetOffsetFromMultiIndex(
src_block_data_multi_id_begin +
Array<index_t, 2>{thread_id_d0, thread_id_d1 * DataPerRead});
mDstMyThreadOffset = DstDesc{}.GetOffsetFromMultiIndex(
dst_block_data_multi_id_begin +
Array<index_t, 2>{thread_id_d0, thread_id_d1 * DataPerRead});
}
__device__ void Run(const Float* __restrict__ p_src, Float* __restrict__ p_dst) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t thread_per_d1 = (L1 + DataPerRead - 1) / DataPerRead;
constexpr index_t thread_per_d0 = BlockSize / thread_per_d1;
constexpr index_t num_active_thread = thread_per_d0 * thread_per_d1;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t src_loop_stride = SrcDesc{}.GetStride(I0) * thread_per_d0;
constexpr index_t dst_loop_stride = DstDesc{}.GetStride(I0) * thread_per_d0;
auto f_copy = [&](index_t iloop) {
*(reinterpret_cast<vector_t*>(p_dst + mDstMyThreadOffset + iloop * dst_loop_stride)) =
*(reinterpret_cast<const vector_t*>(p_src + mSrcMyThreadOffset +
iloop * src_loop_stride));
};
for(index_t iloop = 0; iloop < nloop_d0; ++iloop)
{
f_copy(iloop);
}
constexpr bool has_tail_d0 = (L0 > nloop_d0 * thread_per_d0);
if(has_tail_d0)
{
constexpr index_t tail_d0 = L0 - nloop_d0 * thread_per_d0;
if(get_thread_local_1d_id() < tail_d0 * thread_per_d1)
{
f_copy(nloop_d0);
}
}
}
__device__ constexpr index_t GetRegisterClipboardSize() const
{
static_assert(is_same<Float, float>{}, "wrong! only support float!\n");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t thread_per_d1 = (L1 + DataPerRead - 1) / DataPerRead;
constexpr index_t thread_per_d0 = BlockSize / thread_per_d1;
return DataPerRead * (L0 + thread_per_d0 - 1) / thread_per_d0;
}
__device__ void RunLoadRegisterClipboard(const Float* __restrict__ p_src,
Float* __restrict__ p_clipboard) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t thread_per_d1 = (L1 + DataPerRead - 1) / DataPerRead;
constexpr index_t thread_per_d0 = BlockSize / thread_per_d1;
constexpr index_t num_active_thread = thread_per_d0 * thread_per_d1;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t src_loop_stride = SrcDesc{}.GetStride(I0) * thread_per_d0;
constexpr index_t dst_loop_stride = DstDesc{}.GetStride(I0) * thread_per_d0;
auto f_copy = [&](index_t iloop) {
*(reinterpret_cast<vector_t*>(&p_clipboard[iloop * DataPerRead])) =
*(reinterpret_cast<const vector_t*>(
&p_src[mSrcMyThreadOffset + iloop * src_loop_stride]));
};
for(index_t iloop = 0; iloop < nloop_d0; ++iloop)
{
f_copy(iloop);
}
constexpr bool has_tail_d0 = (L0 > nloop_d0 * thread_per_d0);
if(has_tail_d0)
{
constexpr index_t tail_d0 = L0 - nloop_d0 * thread_per_d0;
if(get_thread_local_1d_id() < tail_d0 * thread_per_d1)
{
f_copy(nloop_d0);
}
}
}
__device__ void RunStoreRegisterClipboard(const Float* __restrict__ p_clipboard,
Float* __restrict__ p_dst) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t thread_per_d1 = (L1 + DataPerRead - 1) / DataPerRead;
constexpr index_t thread_per_d0 = BlockSize / thread_per_d1;
constexpr index_t num_active_thread = thread_per_d0 * thread_per_d1;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t src_loop_stride = SrcDesc{}.GetStride(I0) * thread_per_d0;
constexpr index_t dst_loop_stride = DstDesc{}.GetStride(I0) * thread_per_d0;
auto f_copy = [&](index_t iloop) {
*(reinterpret_cast<vector_t*>(&p_dst[mDstMyThreadOffset + iloop * dst_loop_stride])) =
*(reinterpret_cast<const vector_t*>(&p_clipboard[iloop * DataPerRead]));
};
for(index_t iloop = 0; iloop < nloop_d0; ++iloop)
{
f_copy(iloop);
}
constexpr bool has_tail_d0 = (L0 > nloop_d0 * thread_per_d0);
if(has_tail_d0)
{
constexpr index_t tail_d0 = L0 - nloop_d0 * thread_per_d0;
if(get_thread_local_1d_id() < tail_d0 * thread_per_d1)
{
f_copy(nloop_d0);
}
}
}
#if CK_USE_AMD_INLINE_ASM
__device__ void RunLoadRegisterClipboard_asm(const Float* __restrict__ p_src,
Float* p_clipboard) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t thread_per_d1 = (L1 + DataPerRead - 1) / DataPerRead;
constexpr index_t thread_per_d0 = BlockSize / thread_per_d1;
constexpr index_t num_active_thread = thread_per_d0 * thread_per_d1;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t src_loop_stride = SrcDesc{}.GetStride(I0) * thread_per_d0;
constexpr index_t dst_loop_stride = DstDesc{}.GetStride(I0) * thread_per_d0;
auto f_copy = [&](index_t iloop) {
#if 0
*(reinterpret_cast<vector_t*>(&p_clipboard[iloop * DataPerRead])) =
*(reinterpret_cast<const vector_t*>(&p_src[mSrcMyThreadOffset +
iloop * src_loop_stride]));
#else
static_assert(is_same<float, Float>{} && DataPerRead == 4,
"global_load is only for float4");
global_load(reinterpret_cast<vector_t&>(p_clipboard[iloop * DataPerRead]),
reinterpret_cast<const vector_t*>(
&p_src[mSrcMyThreadOffset + iloop * src_loop_stride]));
#endif
};
for(index_t iloop = 0; iloop < nloop_d0; ++iloop)
{
f_copy(iloop);
}
constexpr bool has_tail_d0 = (L0 > nloop_d0 * thread_per_d0);
if(has_tail_d0)
{
constexpr index_t tail_d0 = L0 - nloop_d0 * thread_per_d0;
if(get_thread_local_1d_id() < tail_d0 * thread_per_d1)
{
f_copy(nloop_d0);
}
}
}
__device__ void RunStoreRegisterClipboard_asm(const Float* __restrict__ p_clipboard,
Float* __restrict__ p_dst) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t thread_per_d1 = (L1 + DataPerRead - 1) / DataPerRead;
constexpr index_t thread_per_d0 = BlockSize / thread_per_d1;
constexpr index_t num_active_thread = thread_per_d0 * thread_per_d1;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t src_loop_stride = SrcDesc{}.GetStride(I0) * thread_per_d0;
constexpr index_t dst_loop_stride = DstDesc{}.GetStride(I0) * thread_per_d0;
auto f_copy = [&](index_t iloop) {
#if 0
*(reinterpret_cast<vector_t*>(&p_dst[mDstMyThreadOffset + iloop * dst_loop_stride]) =
*(reinterpret_cast<const vector_t*>(&p_clipboard[iloop * DataPerRead]);
#else
static_assert(is_same<float, Float>{} && DataPerRead == 4,
"ds_write_b128 is only for float4");
ds_write_b128(reinterpret_cast<const vector_t&>(p_clipboard[iloop * DataPerRead]),
&p_dst[mDstMyThreadOffset + iloop * dst_loop_stride]);
#endif
};
for(index_t iloop = 0; iloop < nloop_d0; ++iloop)
{
f_copy(iloop);
}
constexpr bool has_tail_d0 = (L0 > nloop_d0 * thread_per_d0);
if(has_tail_d0)
{
constexpr index_t tail_d0 = L0 - nloop_d0 * thread_per_d0;
if(get_thread_local_1d_id() < tail_d0 * thread_per_d1)
{
f_copy(nloop_d0);
}
}
}
#endif
};
} // namespace ck
#endif
composable_kernel/include/tensor_operation/blockwise_3d_tensor_op.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_BLOCKWISE_3D_TENSOR_OP_HPP
#define CK_BLOCKWISE_3D_TENSOR_OP_HPP
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
namespace ck {
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class CopyLengths,
index_t DataPerRead>
struct Blockwise3dTensorCopy1
{
using vector_t = typename vector_type<Float, DataPerRead>::MemoryType;
__device__ constexpr Blockwise3dTensorCopy1()
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
static_assert(DataPerRead == 1 ||
(SrcDesc{}.GetStride(I2) == 1 && DstDesc{}.GetStride(I2) == 1),
"wrong! only support stride2 == 1 if DataPerRead > 1!\n");
static_assert(DataPerRead == 1 || DataPerRead == 2 || DataPerRead == 4,
"wrong! only support DataPerRead == 1, 2 or 4!\n");
static_assert(SrcDesc{}.GetStride(I1) % DataPerRead == 0 &&
DstDesc{}.GetStride(I1) % DataPerRead == 0,
"src and dst stride1 should be multiple of DataPerRead to keep alignment");
// we allow out-of-bound read from src in D3 dimension,
// but we need to make sure dst stride2 is big enough,
// so that the out-of-bound write won't contaminate next line in dst
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t read_per_d2 = math::integer_divide_ceil(L2, DataPerRead);
static_assert(read_per_d2 * DataPerRead <= DstDesc{}.GetStride(I1),
"wrong! out-of-bound write will contaminate next line!\n");
}
__device__ void Run(const Float* __restrict__ p_src, Float* __restrict__ p_dst) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto src_desc = SrcDesc{};
constexpr auto dst_desc = DstDesc{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t read_per_d2 = math::integer_divide_ceil(L2, DataPerRead);
constexpr auto ref_desc = make_ConstantTensorDescriptor(Sequence<L0, L1, read_per_d2>{});
constexpr index_t NLoop = ref_desc.GetElementSize() / BlockSize;
auto f_copy = [&](index_t is) {
index_t did[3];
did[0] = is / ref_desc.GetStride(I0);
is -= did[0] * ref_desc.GetStride(I0);
did[1] = is / ref_desc.GetStride(I1);
is -= did[1] * ref_desc.GetStride(I1);
did[2] = is / ref_desc.GetStride(I2);
const index_t src_index =
src_desc.GetOffsetFromMultiIndex(did[0], did[1], did[2] * DataPerRead);
const index_t dst_index =
dst_desc.GetOffsetFromMultiIndex(did[0], did[1], did[2] * DataPerRead);
*(reinterpret_cast<vector_t*>(p_dst + dst_index)) =
*(reinterpret_cast<const vector_t*>(p_src + src_index));
};
for(index_t iloop = 0; iloop < NLoop; ++iloop)
{
index_t is = get_thread_local_1d_id() + iloop * BlockSize;
f_copy(is);
}
constexpr bool has_tail = (ref_desc.GetElementSize() > NLoop * BlockSize);
if(has_tail)
{
index_t is = get_thread_local_1d_id() + NLoop * BlockSize;
if(is < ref_desc.GetElementSize())
{
f_copy(is);
}
}
}
};
// starting point need to be aligned to float4 or float2 or float
// stride3 need to be 1 for both source and destination
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class CopyLengths,
class ThreadPerDims,
index_t DataPerRead>
struct Blockwise3dTensorCopy3
{
using vector_t = typename vector_type<Float, DataPerRead>::MemoryType;
index_t mSrcMyThreadOffset;
index_t mDstMyThreadOffset;
__device__ Blockwise3dTensorCopy3()
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
static_assert(DataPerRead == 1 ||
(SrcDesc{}.GetStride(I2) == 1 && DstDesc{}.GetStride(I2) == 1),
"wrong! only support stride3 == 1 if DataPerRead > 1!\n");
static_assert(DataPerRead == 1 || DataPerRead == 2 || DataPerRead == 4,
"wrong! only support DataPerRead == 1, 2 or 4!\n");
static_assert(
SrcDesc{}.GetStride(I1) % DataPerRead == 0 &&
DstDesc{}.GetStride(I1) % DataPerRead == 0,
"wrong! src and dst stride1 should be multiple of DataPerRead to keep alignment");
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t thread_per_d0 = ThreadPerDims{}.Get(I0);
constexpr index_t thread_per_d1 = ThreadPerDims{}.Get(I1);
constexpr index_t thread_per_d2 = ThreadPerDims{}.Get(I2);
// we allow out-of-bound read from src in D2 dimension,
// but we need to make sure dst stride is big enough,
// so that the out-of-bound write won't contaminate next line in dst
constexpr index_t nloop_d2 = math::integer_divide_ceil(L2, thread_per_d2 * DataPerRead);
static_assert(nloop_d2 * thread_per_d2 * DataPerRead <= DstDesc{}.GetStride(I1),
"wrong! out-of-bound write will contaminate next line!\n");
static_assert(L0 % thread_per_d0 == 0 && L1 % thread_per_d1 == 0,
"wrong! L0, L1, L2 should be divided evenly!\n");
static_assert(BlockSize >= thread_per_d0 * thread_per_d1 * thread_per_d2,
"wrrong! BlockSize is not big enough for ThreadPerDims!");
constexpr index_t num_active_thread =
accumulate_on_sequence(ThreadPerDims{}, math::multiplies<index_t>{}, Number<1>{});
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr auto thread_cluster_desc = make_ConstantTensorDescriptor(ThreadPerDims{});
const auto thread_multi_id =
thread_cluster_desc.GetMultiIndexFrom1dIndex(get_thread_local_1d_id());
mSrcMyThreadOffset = SrcDesc{}.GetOffsetFromMultiIndex(
thread_multi_id[0], thread_multi_id[1], thread_multi_id[2] * DataPerRead);
mDstMyThreadOffset = DstDesc{}.GetOffsetFromMultiIndex(
thread_multi_id[0], thread_multi_id[1], thread_multi_id[2] * DataPerRead);
}
__device__ void Run(const Float* __restrict__ p_src, Float* __restrict__ p_dst) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t thread_per_d0 = ThreadPerDims{}.Get(I0);
constexpr index_t thread_per_d1 = ThreadPerDims{}.Get(I1);
constexpr index_t thread_per_d2 = ThreadPerDims{}.Get(I2);
constexpr index_t num_active_thread = thread_per_d0 * thread_per_d1 * thread_per_d2;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t nloop_d1 = L1 / thread_per_d1;
constexpr index_t nloop_d2 = math::integer_divide_ceil(L2, thread_per_d2 * DataPerRead);
#pragma unroll
for(index_t iloop_d0 = 0; iloop_d0 < nloop_d0; ++iloop_d0)
{
#pragma unroll
for(index_t iloop_d1 = 0; iloop_d1 < nloop_d1; ++iloop_d1)
{
#pragma unroll
for(index_t iloop_d2 = 0; iloop_d2 < nloop_d2; ++iloop_d2)
{
const index_t src_offset =
SrcDesc{}.GetOffsetFromMultiIndex(iloop_d0 * thread_per_d0,
iloop_d1 * thread_per_d1,
iloop_d2 * thread_per_d2 * DataPerRead);
const index_t dst_offset =
DstDesc{}.GetOffsetFromMultiIndex(iloop_d0 * thread_per_d0,
iloop_d1 * thread_per_d1,
iloop_d2 * thread_per_d2 * DataPerRead);
*(reinterpret_cast<vector_t*>(&p_dst[dst_offset + mDstMyThreadOffset])) = *(
reinterpret_cast<const vector_t*>(&p_src[src_offset + mSrcMyThreadOffset]));
}
}
}
}
__device__ static constexpr index_t GetRegisterClipboardSize()
{
static_assert(is_same<Float, float>{}, "wrong! only support float!\n");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t thread_per_d0 = ThreadPerDims{}.Get(I0);
constexpr index_t thread_per_d1 = ThreadPerDims{}.Get(I1);
constexpr index_t thread_per_d2 = ThreadPerDims{}.Get(I2);
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t nloop_d1 = L1 / thread_per_d1;
constexpr index_t nloop_d2 = math::integer_divide_ceil(L2, thread_per_d2 * DataPerRead);
return DataPerRead * nloop_d0 * nloop_d1 * nloop_d2;
}
__device__ void RunLoadRegisterClipboard(const Float* __restrict__ p_src,
Float* __restrict__ p_clipboard) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t thread_per_d0 = ThreadPerDims{}.Get(I0);
constexpr index_t thread_per_d1 = ThreadPerDims{}.Get(I1);
constexpr index_t thread_per_d2 = ThreadPerDims{}.Get(I2);
constexpr index_t num_active_thread = thread_per_d0 * thread_per_d1 * thread_per_d2;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t nloop_d1 = L1 / thread_per_d1;
constexpr index_t nloop_d2 = math::integer_divide_ceil(L2, thread_per_d2 * DataPerRead);
constexpr auto clipboard_desc =
make_ConstantTensorDescriptor(Sequence<nloop_d0, nloop_d1, nloop_d2 * DataPerRead>{});
#pragma unroll
for(index_t iloop_d0 = 0; iloop_d0 < nloop_d0; ++iloop_d0)
{
#pragma unroll
for(index_t iloop_d1 = 0; iloop_d1 < nloop_d1; ++iloop_d1)
{
#pragma unroll
for(index_t iloop_d2 = 0; iloop_d2 < nloop_d2; ++iloop_d2)
{
const index_t src_offset =
SrcDesc{}.GetOffsetFromMultiIndex(iloop_d0 * thread_per_d0,
iloop_d1 * thread_per_d1,
iloop_d2 * thread_per_d2 * DataPerRead);
const index_t clipboard_offset = clipboard_desc.GetOffsetFromMultiIndex(
iloop_d0, iloop_d1, iloop_d2 * DataPerRead);
*(reinterpret_cast<vector_t*>(&p_clipboard[clipboard_offset])) = *(
reinterpret_cast<const vector_t*>(&p_src[src_offset + mSrcMyThreadOffset]));
}
}
}
}
__device__ void RunStoreRegisterClipboard(const Float* __restrict__ p_clipboard,
Float* __restrict__ p_dst) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t thread_per_d0 = ThreadPerDims{}.Get(I0);
constexpr index_t thread_per_d1 = ThreadPerDims{}.Get(I1);
constexpr index_t thread_per_d2 = ThreadPerDims{}.Get(I2);
constexpr index_t num_active_thread = thread_per_d0 * thread_per_d1 * thread_per_d2;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t nloop_d1 = L1 / thread_per_d1;
constexpr index_t nloop_d2 = math::integer_divide_ceil(L2, thread_per_d2 * DataPerRead);
constexpr auto clipboard_desc =
make_ConstantTensorDescriptor(Sequence<nloop_d0, nloop_d1, nloop_d2 * DataPerRead>{});
#pragma unroll
for(index_t iloop_d0 = 0; iloop_d0 < nloop_d0; ++iloop_d0)
{
#pragma unroll
for(index_t iloop_d1 = 0; iloop_d1 < nloop_d1; ++iloop_d1)
{
#pragma unroll
for(index_t iloop_d2 = 0; iloop_d2 < nloop_d2; ++iloop_d2)
{
const index_t clipboard_offset = clipboard_desc.GetOffsetFromMultiIndex(
iloop_d0, iloop_d1, iloop_d2 * DataPerRead);
const index_t dst_offset =
DstDesc{}.GetOffsetFromMultiIndex(iloop_d0 * thread_per_d0,
iloop_d1 * thread_per_d1,
iloop_d2 * thread_per_d2 * DataPerRead);
*(reinterpret_cast<vector_t*>(&p_dst[dst_offset + mDstMyThreadOffset])) =
*(reinterpret_cast<const vector_t*>(&p_clipboard[clipboard_offset]));
}
}
}
}
};
} // namespace ck
#endif
composable_kernel/include/tensor_operation/blockwise_4d_tensor_op.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_BLOCKWISE_4D_TENSOR_OP_HPP
#define CK_BLOCKWISE_4D_TENSOR_OP_HPP
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "threadwise_tensor_slice_copy.hpp"
namespace ck {
template <index_t BlockSize, class Float, class DstDesc, class F>
__device__ void
blockwise_4d_tensor_pointwise_operation_unary(DstDesc, Float* __restrict__ p_dst, F f)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto dst_desc = DstDesc{};
constexpr auto desc = make_ConstantTensorDescriptor_packed(dst_desc.GetLengths());
#if 0
if(get_thread_local_1d_id() == 0)
{
print_ConstantTensorDescriptor(dst_desc, "blockwise_4d_tensor_op_unary: dst_desc: ");
print_ConstantTensorDescriptor(desc, "blockwise_4d_tensor_op_unary: desc: ");
}
#endif
constexpr index_t NLoop = desc.GetElementSize() / BlockSize;
for(index_t iloop = 0; iloop < NLoop; ++iloop)
{
index_t is = get_thread_local_1d_id() + iloop * BlockSize;
const index_t did0 = is / desc.GetStride(I0);
is -= did0 * desc.GetStride(I0);
const index_t did1 = is / desc.GetStride(I1);
is -= did1 * desc.GetStride(I1);
const index_t did2 = is / desc.GetStride(I2);
is -= did2 * desc.GetStride(I2);
const index_t did3 = is / desc.GetStride(I3);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1, did2, did3);
f(p_dst[dindex]);
}
constexpr bool has_tail = (desc.GetElementSize() > NLoop * BlockSize);
if(has_tail)
{
index_t is = get_thread_local_1d_id() + NLoop * BlockSize;
if(is < desc.GetElementSize())
{
const index_t did0 = is / desc.GetStride(I0);
is -= did0 * desc.GetStride(I0);
const index_t did1 = is / desc.GetStride(I1);
is -= did1 * desc.GetStride(I1);
const index_t did2 = is / desc.GetStride(I2);
is -= did2 * desc.GetStride(I2);
const index_t did3 = is / desc.GetStride(I3);
const index_t dindex = dst_desc.GetOffsetFromMultiIndex(did0, did1, did2, did3);
f(p_dst[dindex]);
}
}
}
// Function: p_dst[reorder[i0], reorder[i1], reorder[i2], reorder[i3]] = p_src[i0,i1,i2,i3]
// TODO: in order to optimize mem access for different mem type,
// need to write specialized version
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class SrcOpLengths,
class MapDst2Src,
class F>
__device__ void blockwise_4d_tensor_pointwise_operation_binary_reorder_by_get_dst_from_src(
SrcDesc,
const Float* __restrict__ p_src,
DstDesc,
Float* __restrict__ p_dst,
SrcOpLengths,
MapDst2Src,
F f)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr index_t IR0 = MapDst2Src{}.Get(I0);
constexpr index_t IR1 = MapDst2Src{}.Get(I1);
constexpr index_t IR2 = MapDst2Src{}.Get(I2);
constexpr index_t IR3 = MapDst2Src{}.Get(I3);
constexpr auto src_desc = SrcDesc{};
constexpr auto dst_desc = DstDesc{};
constexpr auto ref_desc = make_ConstantTensorDescriptor_packed(SrcOpLengths{});
constexpr index_t NLoop = ref_desc.GetElementSize() / BlockSize;
for(index_t iloop = 0; iloop < NLoop; ++iloop)
{
index_t is = get_thread_local_1d_id() + iloop * BlockSize;
index_t did[4];
did[0] = is / ref_desc.GetStride(I0);
is -= did[0] * ref_desc.GetStride(I0);
did[1] = is / ref_desc.GetStride(I1);
is -= did[1] * ref_desc.GetStride(I1);
did[2] = is / ref_desc.GetStride(I2);
is -= did[2] * ref_desc.GetStride(I2);
did[3] = is / ref_desc.GetStride(I3);
const index_t src_index = src_desc.GetOffsetFromMultiIndex(did[0], did[1], did[2], did[3]);
const index_t dst_index =
dst_desc.GetOffsetFromMultiIndex(did[IR0], did[IR1], did[IR2], did[IR3]);
f(p_src[src_index], p_dst[dst_index]);
}
constexpr bool has_tail = (ref_desc.GetElementSize() > NLoop * BlockSize);
if(has_tail)
{
index_t is = get_thread_local_1d_id() + NLoop * BlockSize;
if(is < ref_desc.GetElementSize())
{
index_t did[4];
did[0] = is / ref_desc.GetStride(I0);
is -= did[0] * ref_desc.GetStride(I0);
did[1] = is / ref_desc.GetStride(I1);
is -= did[1] * ref_desc.GetStride(I1);
did[2] = is / ref_desc.GetStride(I2);
is -= did[2] * ref_desc.GetStride(I2);
did[3] = is / ref_desc.GetStride(I3);
const index_t src_index =
src_desc.GetOffsetFromMultiIndex(did[0], did[1], did[2], did[3]);
const index_t dst_index =
dst_desc.GetOffsetFromMultiIndex(did[IR0], did[IR1], did[IR2], did[IR3]);
f(p_src[src_index], p_dst[dst_index]);
}
}
}
template <index_t BlockSize, class Float, class DstDesc>
__device__ void blockwise_4d_tensor_set_zero(DstDesc, Float* __restrict__ p_dst)
{
auto f_set_zero = [](Float& v) { v = Float(0); };
blockwise_4d_tensor_pointwise_operation_unary<BlockSize>(DstDesc{}, p_dst, f_set_zero);
}
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class SrcOpLengths,
class MapDst2Src>
__device__ void
blockwise_4d_tensor_copy_reorder_by_get_dst_from_src(SrcDesc,
const Float* __restrict__ p_src,
DstDesc,
Float* __restrict__ p_dst,
SrcOpLengths,
MapDst2Src)
{
auto f_copy = [](const Float& src, Float& dst) { dst = src; };
blockwise_4d_tensor_pointwise_operation_binary_reorder_by_get_dst_from_src<BlockSize>(
SrcDesc{}, p_src, DstDesc{}, p_dst, SrcOpLengths{}, MapDst2Src{}, f_copy);
}
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class CopyLengths,
index_t DataPerRead>
struct Blockwise4dTensorCopy1
{
using vector_t = typename vector_type<Float, DataPerRead>::MemoryType;
__device__ constexpr Blockwise4dTensorCopy1()
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
static_assert(DataPerRead == 1 ||
(SrcDesc{}.GetStride(I3) == 1 && DstDesc{}.GetStride(I3) == 1),
"wrong! only support stride3 == 1 if DataPerRead > 1!\n");
static_assert(DataPerRead == 1 || DataPerRead == 2 || DataPerRead == 4,
"wrong! only support DataPerRead == 1, 2 or 4!\n");
static_assert(SrcDesc{}.GetStride(I2) % DataPerRead == 0 &&
DstDesc{}.GetStride(I2) % DataPerRead == 0,
"src and dst stride2 should be multiple of DataPerRead to keep alignment");
// we allow out-of-bound read from src in D3 dimension,
// but we need to make sure dst stride2 is big enough,
// so that the out-of-bound write won't contaminate next line in dst
constexpr index_t L3 = CopyLengths{}.Get(I3);
constexpr index_t read_per_d3 = math::integer_divide_ceil(L3, DataPerRead);
static_assert(read_per_d3 * DataPerRead <= DstDesc{}.GetStride(I2),
"wrong! out-of-bound write will contaminate next line!\n");
}
__device__ void Run(const Float* __restrict__ p_src, Float* __restrict__ p_dst) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto src_desc = SrcDesc{};
constexpr auto dst_desc = DstDesc{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t L3 = CopyLengths{}.Get(I3);
constexpr index_t read_per_d3 = math::integer_divide_ceil(L3, DataPerRead);
constexpr auto ref_desc =
make_ConstantTensorDescriptor_packed(Sequence<L0, L1, L2, read_per_d3>{});
constexpr index_t NLoop = ref_desc.GetElementSize() / BlockSize;
auto f_copy = [&](index_t is) {
index_t did[4];
did[0] = is / ref_desc.GetStride(I0);
is -= did[0] * ref_desc.GetStride(I0);
did[1] = is / ref_desc.GetStride(I1);
is -= did[1] * ref_desc.GetStride(I1);
did[2] = is / ref_desc.GetStride(I2);
is -= did[2] * ref_desc.GetStride(I2);
did[3] = is / ref_desc.GetStride(I3);
const index_t src_index =
src_desc.GetOffsetFromMultiIndex(did[0], did[1], did[2], did[3] * DataPerRead);
const index_t dst_index =
dst_desc.GetOffsetFromMultiIndex(did[0], did[1], did[2], did[3] * DataPerRead);
*(reinterpret_cast<vector_t*>(p_dst + dst_index)) =
*(reinterpret_cast<const vector_t*>(p_src + src_index));
};
for(index_t iloop = 0; iloop < NLoop; ++iloop)
{
index_t is = get_thread_local_1d_id() + iloop * BlockSize;
f_copy(is);
}
constexpr bool has_tail = (ref_desc.GetElementSize() > NLoop * BlockSize);
if(has_tail)
{
index_t is = get_thread_local_1d_id() + NLoop * BlockSize;
if(is < ref_desc.GetElementSize())
{
f_copy(is);
}
}
}
};
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class DstOpLengths,
class GlobalLowerPads>
struct BlockwiseChwnTensorCopyPadded
{
__device__ void Run(const Float* __restrict__ p_src,
index_t c_block_data_begin,
index_t ho_block_data_begin,
index_t wo_block_data_begin,
index_t n_block_data_begin,
Float* __restrict__ p_dst,
index_t h_block_pad_low,
index_t w_block_pad_low,
index_t h_block_pad_up,
index_t w_block_pad_up) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto src_desc = SrcDesc{};
constexpr auto dst_desc = DstDesc{};
constexpr auto ref_desc = make_ConstantTensorDescriptor_packed(DstOpLengths{});
constexpr auto h_global_pad_low = GlobalLowerPads{}.Get(I0);
constexpr auto w_global_pad_low = GlobalLowerPads{}.Get(I1);
constexpr index_t NLoop = ref_desc.GetElementSize() / BlockSize;
const Float* p_src_tmp = p_src +
src_desc.GetOffsetFromMultiIndex(
c_block_data_begin,
(ho_block_data_begin + h_block_pad_low) - h_global_pad_low,
(wo_block_data_begin + w_block_pad_low) - w_global_pad_low,
n_block_data_begin);
#if 0
if(get_thread_local_1d_id() == 0)
{
print_ConstantTensorDescriptor(src_desc, "src_desc: ");
print_ConstantTensorDescriptor(dst_desc, "dst_desc: ");
print_ConstantTensorDescriptor(ref_desc, "ref_desc: ");
printf("%u %u, \t"
"h_global_pad_low %u w_global_pad_low %u \t"
"h_block_pad_low %u w_block_pad_low %u h_block_pad_up %u w_block_pad_up %u \t"
"\n",
get_block_1d_id(),
get_thread_local_1d_id(),
h_global_pad_low,
w_global_pad_low,
h_block_pad_low,
w_block_pad_low,
h_block_pad_up,
w_block_pad_up);
}
#endif
for(index_t iloop = 0; iloop < NLoop; ++iloop)
{
index_t is = get_thread_local_1d_id() + iloop * BlockSize;
index_t did[4];
did[0] = is / ref_desc.GetStride(I0);
is -= did[0] * ref_desc.GetStride(I0);
did[1] = is / ref_desc.GetStride(I1);
is -= did[1] * ref_desc.GetStride(I1);
did[2] = is / ref_desc.GetStride(I2);
is -= did[2] * ref_desc.GetStride(I2);
did[3] = is / ref_desc.GetStride(I3);
const index_t bindex = dst_desc.GetOffsetFromMultiIndex(did[0], did[1], did[2], did[3]);
p_dst[bindex] =
(did[1] < h_block_pad_low || did[1] + h_block_pad_up >= ref_desc.GetLength(I1) ||
did[2] < w_block_pad_low || did[2] + w_block_pad_up >= ref_desc.GetLength(I2))
? Float(0)
: p_src_tmp[src_desc.GetOffsetFromMultiIndex(did[0], did[1], did[2], did[3])];
}
constexpr bool has_tail = (ref_desc.GetElementSize() > NLoop * BlockSize);
if(has_tail)
{
index_t is = get_thread_local_1d_id() + NLoop * BlockSize;
if(is < ref_desc.GetElementSize())
{
index_t did[4];
did[0] = is / ref_desc.GetStride(I0);
is -= did[0] * ref_desc.GetStride(I0);
did[1] = is / ref_desc.GetStride(I1);
is -= did[1] * ref_desc.GetStride(I1);
did[2] = is / ref_desc.GetStride(I2);
is -= did[2] * ref_desc.GetStride(I2);
did[3] = is / ref_desc.GetStride(I3);
const index_t bindex =
dst_desc.GetOffsetFromMultiIndex(did[0], did[1], did[2], did[3]);
p_dst[bindex] =
(did[1] < h_block_pad_low ||
did[1] + h_block_pad_up >= ref_desc.GetLength(I1) ||
did[2] < w_block_pad_low || did[2] + w_block_pad_up >= ref_desc.GetLength(I2))
? Float(0)
: p_src_tmp[src_desc.GetOffsetFromMultiIndex(
did[0], did[1], did[2], did[3])];
}
}
}
};
// starting point need to be aligned to float4 or float2 or float
// stride3 need to be 1 for both source and destination
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class CopyLengths,
class ThreadPerDims,
index_t DataPerRead>
struct Blockwise4dTensorCopy3
{
using vector_t = typename vector_type<Float, DataPerRead>::MemoryType;
index_t mSrcMyThreadOffset;
index_t mDstMyThreadOffset;
__device__ Blockwise4dTensorCopy3()
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
static_assert(DataPerRead == 1 ||
(SrcDesc{}.GetStride(I3) == 1 && DstDesc{}.GetStride(I3) == 1),
"wrong! only support stride3 == 1 if DataPerRead > 1!\n");
static_assert(DataPerRead == 1 || DataPerRead == 2 || DataPerRead == 4,
"wrong! only support DataPerRead == 1, 2 or 4!\n");
static_assert(
SrcDesc{}.GetStride(I2) % DataPerRead == 0 &&
DstDesc{}.GetStride(I2) % DataPerRead == 0,
"wrong! src and dst stride2 should be multiple of DataPerRead to keep alignment");
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t L3 = CopyLengths{}.Get(I3);
constexpr index_t thread_per_d0 = ThreadPerDims{}.Get(I0);
constexpr index_t thread_per_d1 = ThreadPerDims{}.Get(I1);
constexpr index_t thread_per_d2 = ThreadPerDims{}.Get(I2);
constexpr index_t thread_per_d3 = ThreadPerDims{}.Get(I3);
// we allow out-of-bound read from src in D3 dimension,
// but we need to make sure dst stride is big enough,
// so that the out-of-bound write won't contaminate next line in dst
constexpr index_t nloop_d3 = math::integer_divide_ceil(L3, thread_per_d3 * DataPerRead);
static_assert(nloop_d3 * thread_per_d3 * DataPerRead <= DstDesc{}.GetStride(I2),
"wrong! out-of-bound write will contaminate next line!\n");
static_assert(L0 % thread_per_d0 == 0 && L1 % thread_per_d1 == 0 && L2 % thread_per_d2 == 0,
"wrong! L0, L1, L2 should be divided evenly!\n");
static_assert(BlockSize >= thread_per_d0 * thread_per_d1 * thread_per_d2 * thread_per_d3,
"wrrong! BlockSize is not big enough for ThreadPerDims!");
constexpr index_t num_active_thread =
accumulate_on_sequence(ThreadPerDims{}, math::multiplies<index_t>{}, Number<1>{});
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr auto thread_cluster_desc = make_ConstantTensorDescriptor_packed(ThreadPerDims{});
const auto thread_multi_id =
thread_cluster_desc.GetMultiIndexFrom1dIndex(get_thread_local_1d_id());
mSrcMyThreadOffset = SrcDesc{}.GetOffsetFromMultiIndex(thread_multi_id[0],
thread_multi_id[1],
thread_multi_id[2],
thread_multi_id[3] * DataPerRead);
mDstMyThreadOffset = DstDesc{}.GetOffsetFromMultiIndex(thread_multi_id[0],
thread_multi_id[1],
thread_multi_id[2],
thread_multi_id[3] * DataPerRead);
}
__device__ void Run(const Float* __restrict__ p_src, Float* __restrict__ p_dst) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t L3 = CopyLengths{}.Get(I3);
constexpr index_t thread_per_d0 = ThreadPerDims{}.Get(I0);
constexpr index_t thread_per_d1 = ThreadPerDims{}.Get(I1);
constexpr index_t thread_per_d2 = ThreadPerDims{}.Get(I2);
constexpr index_t thread_per_d3 = ThreadPerDims{}.Get(I3);
constexpr index_t num_active_thread =
thread_per_d0 * thread_per_d1 * thread_per_d2 * thread_per_d3;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t nloop_d1 = L1 / thread_per_d1;
constexpr index_t nloop_d2 = L2 / thread_per_d2;
constexpr index_t nloop_d3 = math::integer_divide_ceil(L3, thread_per_d3 * DataPerRead);
#pragma unroll
for(index_t iloop_d0 = 0; iloop_d0 < nloop_d0; ++iloop_d0)
{
#pragma unroll
for(index_t iloop_d1 = 0; iloop_d1 < nloop_d1; ++iloop_d1)
{
#pragma unroll
for(index_t iloop_d2 = 0; iloop_d2 < nloop_d2; ++iloop_d2)
{
#pragma unroll
for(index_t iloop_d3 = 0; iloop_d3 < nloop_d3; ++iloop_d3)
{
const index_t src_offset = SrcDesc{}.GetOffsetFromMultiIndex(
iloop_d0 * thread_per_d0,
iloop_d1 * thread_per_d1,
iloop_d2 * thread_per_d2,
iloop_d3 * thread_per_d3 * DataPerRead);
const index_t dst_offset = DstDesc{}.GetOffsetFromMultiIndex(
iloop_d0 * thread_per_d0,
iloop_d1 * thread_per_d1,
iloop_d2 * thread_per_d2,
iloop_d3 * thread_per_d3 * DataPerRead);
*(reinterpret_cast<vector_t*>(&p_dst[dst_offset + mDstMyThreadOffset])) =
*(reinterpret_cast<const vector_t*>(
&p_src[src_offset + mSrcMyThreadOffset]));
}
}
}
}
}
__device__ constexpr index_t GetRegisterClipboardSize() const
{
static_assert(is_same<Float, float>{}, "wrong! only support float!\n");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t L3 = CopyLengths{}.Get(I3);
constexpr index_t thread_per_d0 = ThreadPerDims{}.Get(I0);
constexpr index_t thread_per_d1 = ThreadPerDims{}.Get(I1);
constexpr index_t thread_per_d2 = ThreadPerDims{}.Get(I2);
constexpr index_t thread_per_d3 = ThreadPerDims{}.Get(I3);
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t nloop_d1 = L1 / thread_per_d1;
constexpr index_t nloop_d2 = L2 / thread_per_d2;
constexpr index_t nloop_d3 = math::integer_divide_ceil(L3, thread_per_d3 * DataPerRead);
return DataPerRead * nloop_d0 * nloop_d1 * nloop_d2 * nloop_d3;
}
__device__ void RunLoadRegisterClipboard(const Float* __restrict__ p_src,
Float* __restrict__ p_clipboard) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t L3 = CopyLengths{}.Get(I3);
constexpr index_t thread_per_d0 = ThreadPerDims{}.Get(I0);
constexpr index_t thread_per_d1 = ThreadPerDims{}.Get(I1);
constexpr index_t thread_per_d2 = ThreadPerDims{}.Get(I2);
constexpr index_t thread_per_d3 = ThreadPerDims{}.Get(I3);
constexpr index_t num_active_thread =
thread_per_d0 * thread_per_d1 * thread_per_d2 * thread_per_d3;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t nloop_d1 = L1 / thread_per_d1;
constexpr index_t nloop_d2 = L2 / thread_per_d2;
constexpr index_t nloop_d3 = math::integer_divide_ceil(L3, thread_per_d3 * DataPerRead);
constexpr auto clipboard_desc = make_ConstantTensorDescriptor_packed(
Sequence<nloop_d0, nloop_d1, nloop_d2, nloop_d3 * DataPerRead>{});
#pragma unroll
for(index_t iloop_d0 = 0; iloop_d0 < nloop_d0; ++iloop_d0)
{
#pragma unroll
for(index_t iloop_d1 = 0; iloop_d1 < nloop_d1; ++iloop_d1)
{
#pragma unroll
for(index_t iloop_d2 = 0; iloop_d2 < nloop_d2; ++iloop_d2)
{
#pragma unroll
for(index_t iloop_d3 = 0; iloop_d3 < nloop_d3; ++iloop_d3)
{
const index_t src_offset = SrcDesc{}.GetOffsetFromMultiIndex(
iloop_d0 * thread_per_d0,
iloop_d1 * thread_per_d1,
iloop_d2 * thread_per_d2,
iloop_d3 * thread_per_d3 * DataPerRead);
const index_t clipboard_offset = clipboard_desc.GetOffsetFromMultiIndex(
iloop_d0, iloop_d1, iloop_d2, iloop_d3 * DataPerRead);
*(reinterpret_cast<vector_t*>(&p_clipboard[clipboard_offset])) =
*(reinterpret_cast<const vector_t*>(
&p_src[src_offset + mSrcMyThreadOffset]));
}
}
}
}
}
__device__ void RunStoreRegisterClipboard(const Float* __restrict__ p_clipboard,
Float* __restrict__ p_dst) const
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr index_t L0 = CopyLengths{}.Get(I0);
constexpr index_t L1 = CopyLengths{}.Get(I1);
constexpr index_t L2 = CopyLengths{}.Get(I2);
constexpr index_t L3 = CopyLengths{}.Get(I3);
constexpr index_t thread_per_d0 = ThreadPerDims{}.Get(I0);
constexpr index_t thread_per_d1 = ThreadPerDims{}.Get(I1);
constexpr index_t thread_per_d2 = ThreadPerDims{}.Get(I2);
constexpr index_t thread_per_d3 = ThreadPerDims{}.Get(I3);
constexpr index_t num_active_thread =
thread_per_d0 * thread_per_d1 * thread_per_d2 * thread_per_d3;
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
constexpr index_t nloop_d0 = L0 / thread_per_d0;
constexpr index_t nloop_d1 = L1 / thread_per_d1;
constexpr index_t nloop_d2 = L2 / thread_per_d2;
constexpr index_t nloop_d3 = math::integer_divide_ceil(L3, thread_per_d3 * DataPerRead);
constexpr auto clipboard_desc = make_ConstantTensorDescriptor_packed(
Sequence<nloop_d0, nloop_d1, nloop_d2, nloop_d3 * DataPerRead>{});
#pragma unroll
for(index_t iloop_d0 = 0; iloop_d0 < nloop_d0; ++iloop_d0)
{
#pragma unroll
for(index_t iloop_d1 = 0; iloop_d1 < nloop_d1; ++iloop_d1)
{
#pragma unroll
for(index_t iloop_d2 = 0; iloop_d2 < nloop_d2; ++iloop_d2)
{
#pragma unroll
for(index_t iloop_d3 = 0; iloop_d3 < nloop_d3; ++iloop_d3)
{
const index_t clipboard_offset = clipboard_desc.GetOffsetFromMultiIndex(
iloop_d0, iloop_d1, iloop_d2, iloop_d3 * DataPerRead);
const index_t dst_offset = DstDesc{}.GetOffsetFromMultiIndex(
iloop_d0 * thread_per_d0,
iloop_d1 * thread_per_d1,
iloop_d2 * thread_per_d2,
iloop_d3 * thread_per_d3 * DataPerRead);
*(reinterpret_cast<vector_t*>(&p_dst[dst_offset + mDstMyThreadOffset])) =
*(reinterpret_cast<const vector_t*>(&p_clipboard[clipboard_offset]));
}
}
}
}
}
};
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class SrcOpLengths,
class MapDst2Src>
struct Blockwise4dTensorCopyReorder1
{
__device__ void Run(const Float* __restrict__ p_src, Float* __restrict__ p_dst) const
{
auto f_copy = [](const Float& src, Float& dst) { dst = src; };
blockwise_4d_tensor_pointwise_operation_binary_reorder_by_get_dst_from_src<BlockSize>(
SrcDesc{}, p_src, DstDesc{}, p_dst, SrcOpLengths{}, MapDst2Src{}, f_copy);
}
};
} // namespace
#endif
composable_kernel/include/tensor_operation/blockwise_batched_gemm.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_BLOCKWISE_BATCHED_GEMM_HPP
#define CK_BLOCKWISE_BATCHED_GEMM_HPP
#include "common_header.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "threadwise_gemm.hpp"
namespace ck {
template <index_t BlockSize,
class BlockMatrixA,
class BlockMatrixB,
class ThreadMatrixC,
index_t BlockMatrixStrideA,
index_t BlockMatrixStrideB,
index_t ThreadMatrixStrideC,
index_t BatchSize,
index_t MPerThreadSubC,
index_t NPerThreadSubC,
index_t MLevel0Cluster,
index_t NLevel0Cluster,
index_t MLevel1Cluster,
index_t NLevel1Cluster,
index_t KPerThreadLoop,
index_t BatchPerThread,
index_t DataPerReadA,
index_t DataPerReadB>
struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
{
index_t mMyThreadOffsetA = 0;
index_t mMyThreadOffsetB = 0;
struct MatrixIndex
{
index_t batch;
index_t row;
index_t col;
};
__device__ BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2()
{
static_assert(BatchSize % BatchPerThread == 0,
"wrong! BatchSize is not dividable by BatchPerThread");
constexpr index_t BatchThreadWork = BatchSize / BatchPerThread;
constexpr index_t ThreadPerLevel1Cluster =
MLevel0Cluster * NLevel0Cluster * MLevel1Cluster * NLevel1Cluster;
static_assert(BlockSize == BatchThreadWork * ThreadPerLevel1Cluster,
"wrong! wrong blocksize\n");
constexpr auto a_block_mtx = BlockMatrixA{};
constexpr auto b_block_mtx = BlockMatrixB{};
constexpr auto c_thread_mtx = ThreadMatrixC{};
static_assert(a_block_mtx.NRow() == b_block_mtx.NRow(),
"wrong! K dimension not consistent\n");
constexpr index_t M = a_block_mtx.NCol(); // A is transposed
constexpr index_t N = b_block_mtx.NCol();
constexpr index_t MPerThread = c_thread_mtx.NRow();
constexpr index_t NPerThread = c_thread_mtx.NCol();
static_assert((MPerThread % MPerThreadSubC == 0) && (NPerThread % NPerThreadSubC == 0),
"wrong! Cannot evenly divide thread work among repeat \n");
constexpr index_t MRepeat = MPerThread / MPerThreadSubC;
constexpr index_t NRepeat = NPerThread / NPerThreadSubC;
static_assert((M % MRepeat == 0) && (N % NRepeat == 0),
"wrong! Cannot evenly divide work among repeat\n");
constexpr index_t MPerLevel1Cluster = M / MRepeat;
constexpr index_t NPerLevel1Cluster = N / NRepeat;
static_assert((MPerLevel1Cluster % MLevel1Cluster == 0) &&
(NPerLevel1Cluster % NLevel1Cluster == 0),
"wrong! Cannot evenly divide work among Level1Cluster\n");
constexpr index_t MPerLevel0Cluster = MPerLevel1Cluster / MLevel1Cluster;
constexpr index_t NPerLevel0Cluster = NPerLevel1Cluster / NLevel1Cluster;
static_assert((MPerLevel0Cluster % MLevel0Cluster == 0) &&
(NPerLevel0Cluster % NLevel0Cluster == 0),
"wrong! Cannot evenly divide work among Level0Cluster\n");
static_assert((MPerThreadSubC == MPerLevel0Cluster / MLevel0Cluster) &&
(NPerThreadSubC == NPerLevel0Cluster / NLevel0Cluster),
"wrong! thread work size is wrong\n");
const auto c_thread_mtx_index = GetBeginOfThreadMatrixC(get_thread_local_1d_id());
mMyThreadOffsetA = c_thread_mtx_index.batch * BlockMatrixStrideA +
a_block_mtx.GetOffsetFromMultiIndex(0, c_thread_mtx_index.row);
mMyThreadOffsetB = c_thread_mtx_index.batch * BlockMatrixStrideB +
b_block_mtx.GetOffsetFromMultiIndex(0, c_thread_mtx_index.col);
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantMatrixDescriptor(BlockMatrixA{}, "a_block_mtx: ");
print_ConstantMatrixDescriptor(BlockMatrixB{}, "b_block_mtx: ");
print_ConstantMatrixDescriptor(ThreadMatrixC{}, "c_thread_mtx: ");
printf("%u %u, %u %u %u, %u %u\n",
get_block_1d_id(),
get_thread_local_1d_id(),
c_thread_mtx_index.batch,
c_thread_mtx_index.row,
c_thread_mtx_index.col,
mMyThreadOffsetA,
mMyThreadOffsetB);
}
#endif
}
__device__ MatrixIndex GetBeginOfThreadMatrixC(index_t thread_id) const
{
constexpr index_t ThreadPerLevel1Cluster =
MLevel0Cluster * NLevel0Cluster * MLevel1Cluster * NLevel1Cluster;
constexpr index_t ThreadPerLevel0Cluster = MLevel0Cluster * NLevel0Cluster;
index_t batch_work_id = thread_id / ThreadPerLevel1Cluster;
index_t cluster_id = thread_id - batch_work_id * ThreadPerLevel1Cluster;
index_t level1_id = cluster_id / ThreadPerLevel0Cluster;
index_t level1_m_id = level1_id / NLevel1Cluster;
index_t level1_n_id = level1_id % NLevel1Cluster;
index_t level0_id = cluster_id % ThreadPerLevel0Cluster;
index_t level0_m_id = level0_id / NLevel0Cluster;
index_t level0_n_id = level0_id % NLevel0Cluster;
constexpr index_t MPerLevel0Cluster = MPerThreadSubC * MLevel0Cluster;
constexpr index_t NPerLevel0Cluster = NPerThreadSubC * NLevel0Cluster;
return MatrixIndex{batch_work_id * BatchPerThread,
level1_m_id * MPerLevel0Cluster + level0_m_id * MPerThreadSubC,
level1_n_id * NPerLevel0Cluster + level0_n_id * NPerThreadSubC};
}
// this should be optimized away because input will be known at compile time
__device__ static MatrixIndex
GetDistanceFromBeginOfThreadMatrixC(index_t batch_in_c, index_t m_in_c, index_t n_in_c)
{
constexpr auto c_thread_mtx = ThreadMatrixC{};
constexpr index_t MPerThread = c_thread_mtx.NRow();
constexpr index_t NPerThread = c_thread_mtx.NCol();
constexpr index_t MRepeat = MPerThread / MPerThreadSubC;
constexpr index_t NRepeat = NPerThread / NPerThreadSubC;
constexpr index_t MPerLevel1Cluster = MPerThreadSubC * MLevel0Cluster * MLevel1Cluster;
constexpr index_t NPerLevel1Cluster = NPerThreadSubC * NLevel0Cluster * NLevel1Cluster;
index_t m_repeat = m_in_c / MPerThreadSubC;
index_t n_repeat = n_in_c / NPerThreadSubC;
index_t m_in_sub_c = m_in_c % MPerThreadSubC;
index_t n_in_sub_c = n_in_c % NPerThreadSubC;
return MatrixIndex{batch_in_c,
m_repeat * MPerLevel1Cluster + m_in_sub_c,
n_repeat * NPerLevel1Cluster + n_in_sub_c};
}
template <class FloatA, class FloatB, class FloatC>
__device__ void Run(const FloatA* __restrict__ p_a_block,
const FloatB* __restrict__ p_b_block,
FloatC* __restrict__ p_c_thread) const
{
constexpr auto True = integral_constant<bool, true>{};
constexpr auto False = integral_constant<bool, false>{};
constexpr auto a_block_mtx = BlockMatrixA{};
constexpr auto b_block_mtx = BlockMatrixB{};
constexpr auto c_thread_mtx = ThreadMatrixC{};
constexpr index_t KPerBlock = a_block_mtx.NRow(); // A is transposed
constexpr index_t MPerThread = c_thread_mtx.NRow();
constexpr index_t NPerThread = c_thread_mtx.NCol();
// thread A, B for GEMM
// A is transposed, b is not
constexpr auto a_thread_mtx =
make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<MPerThread>{});
constexpr auto b_thread_mtx =
make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<NPerThread>{});
// thread A-sub, B-sub for copy
constexpr auto a_thread_sub_mtx = make_ConstantMatrixDescriptor(
Number<KPerThreadLoop>{}, Number<MPerThreadSubC>{}, Number<MPerThread>{});
constexpr auto b_thread_sub_mtx = make_ConstantMatrixDescriptor(
Number<KPerThreadLoop>{}, Number<NPerThreadSubC>{}, Number<NPerThread>{});
FloatA p_a_thread[a_thread_mtx.GetElementSpace()];
FloatB p_b_thread[b_thread_mtx.GetElementSpace()];
constexpr index_t MPerLevel1Cluster = MPerThreadSubC * MLevel0Cluster * MLevel1Cluster;
constexpr index_t NPerLevel1Cluster = NPerThreadSubC * NLevel0Cluster * NLevel1Cluster;
constexpr index_t MRepeat = MPerThread / MPerThreadSubC;
constexpr index_t NRepeat = NPerThread / NPerThreadSubC;
// loop over k
#pragma unroll
for(index_t k_begin = 0; k_begin < KPerBlock; k_begin += KPerThreadLoop)
{
// loop over batch
#pragma unroll
for(index_t ib = 0; ib < BatchPerThread; ++ib)
{
// read next batch of a, b
if(BlockMatrixStrideA != 0 or ib == 0)
{
#pragma unroll
for(index_t m_repeat = 0; m_repeat < MRepeat; ++m_repeat)
{
threadwise_matrix_copy(
a_block_mtx,
p_a_block +
a_block_mtx.GetOffsetFromMultiIndex(k_begin,
m_repeat * MPerLevel1Cluster) +
ib * BlockMatrixStrideA + mMyThreadOffsetA,
a_thread_mtx,
p_a_thread +
a_thread_mtx.GetOffsetFromMultiIndex(0, m_repeat * MPerThreadSubC),
a_thread_sub_mtx.GetLengths(),
Number<DataPerReadA>{});
}
}
if(BlockMatrixStrideB != 0 or ib == 0)
{
#pragma unroll
for(index_t n_repeat = 0; n_repeat < NRepeat; ++n_repeat)
{
threadwise_matrix_copy(
b_block_mtx,
p_b_block +
b_block_mtx.GetOffsetFromMultiIndex(k_begin,
n_repeat * NPerLevel1Cluster) +
ib * BlockMatrixStrideB + mMyThreadOffsetB,
b_thread_mtx,
p_b_thread +
b_thread_mtx.GetOffsetFromMultiIndex(0, n_repeat * NPerThreadSubC),
b_thread_sub_mtx.GetLengths(),
Number<DataPerReadB>{});
}
}
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
printf("a: %f %f %f %f %f %f %f %f, b: %f %f %f %f %f %f %f %f\n",
p_a_thread[0],
p_a_thread[1],
p_a_thread[2],
p_a_thread[3],
p_a_thread[4],
p_a_thread[5],
p_a_thread[6],
p_a_thread[7],
p_b_thread[0],
p_b_thread[1],
p_b_thread[2],
p_b_thread[3],
p_b_thread[4],
p_b_thread[5],
p_b_thread[6],
p_b_thread[7]);
}
#endif
threadwise_gemm(a_thread_mtx,
True,
p_a_thread,
b_thread_mtx,
False,
p_b_thread,
c_thread_mtx,
False,
p_c_thread + ib * ThreadMatrixStrideC);
}
}
}
#if CK_USE_AMD_INLINE_ASM
template <class FloatA, class FloatB, class FloatC>
__device__ void Run_amd_asm(const FloatA* __restrict__ p_a_block,
const FloatB* __restrict__ p_b_block,
FloatC* __restrict__ p_c_thread) const
{
constexpr auto a_block_mtx = BlockMatrixA{};
constexpr auto b_block_mtx = BlockMatrixB{};
constexpr auto c_thread_mtx = ThreadMatrixC{};
constexpr index_t K = a_block_mtx.NRow(); // A is transposed
constexpr index_t MPerThread = c_thread_mtx.NRow();
constexpr index_t NPerThread = c_thread_mtx.NCol();
// thread A, B for GEMM
// A is transposed, b is not
constexpr auto a_thread_mtx =
make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<MPerThread>{});
constexpr auto b_thread_mtx =
make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<NPerThread>{});
// thread A-sub, B-sub for copy
constexpr auto a_thread_sub_mtx = make_ConstantMatrixDescriptor(
Number<KPerThreadLoop>{}, Number<MPerThreadSubC>{}, Number<MPerThread>{});
constexpr auto b_thread_sub_mtx = make_ConstantMatrixDescriptor(
Number<KPerThreadLoop>{}, Number<NPerThreadSubC>{}, Number<NPerThread>{});
FloatA p_a_thread[a_thread_mtx.GetElementSpace()];
FloatB p_b_thread[b_thread_mtx.GetElementSpace()];
constexpr index_t MPerLevel1Cluster = MPerThreadSubC * MLevel0Cluster * MLevel1Cluster;
constexpr index_t NPerLevel1Cluster = NPerThreadSubC * NLevel0Cluster * NLevel1Cluster;
// assertion for inline asm
static_assert(is_same<FloatA, float>{} && is_same<FloatB, float>{} &&
is_same<FloatC, float>{},
"Run_amd_asm only deal with float\n");
static_assert(MPerThreadSubC == 4 && NPerThreadSubC == 4 && KPerThreadLoop == 1 &&
MPerThread == 8 && NPerThread == 8,
"Run_amd_asm cannot deal with this GEMM shape yet\n");
static_assert(DataPerReadA == 4 && DataPerReadB == 4, "Run_amd_asm only do float4 read\n");
static_assert(BlockMatrixStrideA == 0 && BatchPerThread == 1,
"Run_amd_asm can only deal with BlockMatrixStrideA == 0 && BatchPerThread == "
"1 for now\n");
using Float4 = vector_type<float, 4>::MemoryType;
Float4* reg_a = (Float4*)(p_a_thread);
Float4* reg_b = (Float4*)(p_b_thread);
Float4* reg_c = (Float4*)(p_c_thread);
reg_a[0] = *reinterpret_cast<const Float4*>(&p_a_block[mMyThreadOffsetA]);
reg_b[0] = *reinterpret_cast<const Float4*>(&p_b_block[mMyThreadOffsetB]);
reg_b[1] = *reinterpret_cast<const Float4*>(
&p_b_block[b_block_mtx.GetOffsetFromMultiIndex(0, NPerLevel1Cluster) +
mMyThreadOffsetB]);
reg_a[1] = *reinterpret_cast<const Float4*>(
&p_a_block[a_block_mtx.GetOffsetFromMultiIndex(0, MPerLevel1Cluster) +
mMyThreadOffsetA]);
outerProduct4x4(reg_a[0], reg_b[0], reg_c[0], reg_c[2], reg_c[4], reg_c[6]);
outerProduct4x4(reg_a[0], reg_b[1], reg_c[1], reg_c[3], reg_c[5], reg_c[7]);
#pragma unroll
for(index_t k = 1; k < K; ++k)
{
reg_a[0] = *reinterpret_cast<const Float4*>(
&p_a_block[a_block_mtx.GetOffsetFromMultiIndex(k, 0) + mMyThreadOffsetA]);
outerProduct4x4(reg_a[1], reg_b[0], reg_c[8], reg_c[10], reg_c[12], reg_c[14]);
reg_b[0] = *reinterpret_cast<const Float4*>(
&p_b_block[b_block_mtx.GetOffsetFromMultiIndex(k, 0) + mMyThreadOffsetB]);
outerProduct4x4(reg_a[1], reg_b[1], reg_c[9], reg_c[11], reg_c[13], reg_c[15]);
reg_b[1] = *reinterpret_cast<const Float4*>(
&p_b_block[b_block_mtx.GetOffsetFromMultiIndex(k, NPerLevel1Cluster) +
mMyThreadOffsetB]);
reg_a[1] = *reinterpret_cast<const Float4*>(
&p_a_block[a_block_mtx.GetOffsetFromMultiIndex(k, MPerLevel1Cluster) +
mMyThreadOffsetA]);
outerProduct4x4(reg_a[0], reg_b[0], reg_c[0], reg_c[2], reg_c[4], reg_c[6]);
outerProduct4x4(reg_a[0], reg_b[1], reg_c[1], reg_c[3], reg_c[5], reg_c[7]);
}
outerProduct4x4(reg_a[1], reg_b[0], reg_c[8], reg_c[10], reg_c[12], reg_c[14]);
outerProduct4x4(reg_a[1], reg_b[1], reg_c[9], reg_c[11], reg_c[13], reg_c[15]);
}
template <class FloatA, class FloatB, class FloatC>
__device__ void Run_asm_v2(const FloatA* __restrict__ p_a_block,
const FloatB* __restrict__ p_b_block,
FloatC* __restrict__ p_c_thread) const
{
constexpr auto a_block_mtx = BlockMatrixA{};
constexpr auto b_block_mtx = BlockMatrixB{};
constexpr auto c_thread_mtx = ThreadMatrixC{};
constexpr index_t M = a_block_mtx.NCol();
constexpr index_t N = b_block_mtx.NCol();
constexpr index_t K = a_block_mtx.NRow(); // A is transposed
constexpr index_t MPerThread = c_thread_mtx.NRow();
constexpr index_t NPerThread = c_thread_mtx.NCol();
// thread A, B for GEMM
// A is transposed, b is not
constexpr auto a_thread_mtx =
make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<MPerThread>{});
constexpr auto b_thread_mtx =
make_ConstantMatrixDescriptor(Number<KPerThreadLoop>{}, Number<NPerThread>{});
// thread A-sub, B-sub for copy
constexpr auto a_thread_sub_mtx = make_ConstantMatrixDescriptor(
Number<KPerThreadLoop>{}, Number<MPerThreadSubC>{}, Number<MPerThread>{});
constexpr auto b_thread_sub_mtx = make_ConstantMatrixDescriptor(
Number<KPerThreadLoop>{}, Number<NPerThreadSubC>{}, Number<NPerThread>{});
FloatA p_a_thread[a_thread_mtx.GetElementSpace()];
FloatB p_b_thread[b_thread_mtx.GetElementSpace()];
constexpr index_t MPerLevel1Cluster = MPerThreadSubC * MLevel0Cluster * MLevel1Cluster;
constexpr index_t NPerLevel1Cluster = NPerThreadSubC * NLevel0Cluster * NLevel1Cluster;
// assertion for inline asm
static_assert(is_same<FloatA, float>{} && is_same<FloatB, float>{} &&
is_same<FloatC, float>{},
"Run_amd_asm only deal with float\n");
static_assert(MPerThreadSubC == 4 && NPerThreadSubC == 4 && KPerThreadLoop == 1 &&
MPerThread == 8 && NPerThread == 8,
"Run_amd_asm cannot deal with this GEMM shape yet\n");
static_assert(DataPerReadA == 4 && DataPerReadB == 4, "Run_amd_asm only do float4 read\n");
static_assert(BlockMatrixStrideA == 0 && BatchPerThread == 1,
"Run_amd_asm can only deal with BlockMatrixStrideA == 0 && BatchPerThread == "
"1 for now\n");
using Float4 = vector_type<float, 4>::MemoryType;
Float4* reg_a = (Float4*)(p_a_thread);
Float4* reg_b = (Float4*)(p_b_thread);
Float4* reg_c = (Float4*)(p_c_thread);
void* a_lds_loc = (void*)(p_a_block + mMyThreadOffsetA);
void* b_lds_loc = (void*)(p_b_block + mMyThreadOffsetB);
constexpr index_t a_lds_row_stride = sizeof(float) * a_block_mtx.RowStride();
constexpr index_t b_lds_row_stride = sizeof(float) * b_block_mtx.RowStride();
constexpr index_t a_lds_cluster_col_stride = sizeof(float) * MPerLevel1Cluster;
constexpr index_t b_lds_cluster_col_stride = sizeof(float) * NPerLevel1Cluster;
ds_read_b128(reg_a[0], a_lds_loc, 0);
ds_read_b128(reg_b[0], b_lds_loc, 0);
ds_read_b128(reg_b[1], b_lds_loc, b_lds_cluster_col_stride);
ds_read_b128(reg_a[1], a_lds_loc, a_lds_cluster_col_stride);
lgkmcnt(2);
outerProduct4x4(reg_a[0], reg_b[0], reg_c[0], reg_c[2], reg_c[4], reg_c[6]);
lgkmcnt(1);
outerProduct4x4(reg_a[0], reg_b[1], reg_c[1], reg_c[3], reg_c[5], reg_c[7]);
#pragma unroll
for(index_t k = 1; k < K; ++k)
{
ds_read_b128(reg_a[0], a_lds_loc, k * a_lds_row_stride);
lgkmcnt(1);
outerProduct4x4(reg_a[1], reg_b[0], reg_c[8], reg_c[10], reg_c[12], reg_c[14]);
ds_read_b128(reg_b[0], b_lds_loc, k * b_lds_row_stride);
outerProduct4x4(reg_a[1], reg_b[1], reg_c[9], reg_c[11], reg_c[13], reg_c[15]);
ds_read_b128(reg_b[1], b_lds_loc, b_lds_cluster_col_stride + k * b_lds_row_stride);
ds_read_b128(reg_a[1], a_lds_loc, a_lds_cluster_col_stride + k * a_lds_row_stride);
lgkmcnt(2);
outerProduct4x4(reg_a[0], reg_b[0], reg_c[0], reg_c[2], reg_c[4], reg_c[6]);
lgkmcnt(1);
outerProduct4x4(reg_a[0], reg_b[1], reg_c[1], reg_c[3], reg_c[5], reg_c[7]);
}
lgkmcnt(0);
outerProduct4x4(reg_a[1], reg_b[0], reg_c[8], reg_c[10], reg_c[12], reg_c[14]);
outerProduct4x4(reg_a[1], reg_b[1], reg_c[9], reg_c[11], reg_c[13], reg_c[15]);
}
#endif
template <class BlockMatrixC, index_t BlockMatrixStrideC, class FloatC>
__device__ void CopyThreadMatrixCToBlockMatrixC(const FloatC* __restrict__ p_c_thread,
FloatC* __restrict__ p_c_block) const
{
constexpr auto c_block_mtx = BlockMatrixC{};
constexpr auto c_thread_mtx = ThreadMatrixC{};
constexpr index_t MPerThread = c_thread_mtx.NRow();
constexpr index_t NPerThread = c_thread_mtx.NCol();
constexpr auto c_thread_sub_mtx = make_ConstantMatrixDescriptor(
Number<MPerThreadSubC>{}, Number<NPerThreadSubC>{}, Number<NPerThread>{});
constexpr index_t MPerLevel1Cluster = MPerThreadSubC * MLevel0Cluster * MLevel1Cluster;
constexpr index_t NPerLevel1Cluster = NPerThreadSubC * NLevel0Cluster * NLevel1Cluster;
constexpr index_t MRepeat = MPerThread / MPerThreadSubC;
constexpr index_t NRepeat = NPerThread / NPerThreadSubC;
const auto c_thread_mtx_begin = GetBeginOfThreadMatrixC(get_thread_local_1d_id());
const index_t c_thread_offset =
c_thread_mtx_begin.batch * BlockMatrixStrideC +
c_block_mtx.GetOffsetFromMultiIndex(c_thread_mtx_begin.row, c_thread_mtx_begin.col);
for(index_t m_repeat = 0; m_repeat < MRepeat; ++m_repeat)
{
for(index_t n_repeat = 0; n_repeat < NRepeat; ++n_repeat)
{
threadwise_matrix_copy(
c_thread_sub_mtx,
p_c_thread +
c_thread_sub_mtx.GetOffsetFromMultiIndex(m_repeat * MPerLevel1Cluster,
n_repeat * NPerLevel1Cluster),
c_block_mtx,
p_c_block +
c_block_mtx.GetOffsetFromMultiIndex(m_repeat * MPerLevel1Cluster,
n_repeat * NPerLevel1Cluster) +
c_thread_offset,
c_thread_sub_mtx.GetLengths());
}
}
}
};
} // namespace
#endif
composable_kernel/include/tensor_operation/blockwise_tensor_slice_copy.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_BLOCKWISE_TENSOR_SLICE_COPY_HPP
#define CK_BLOCKWISE_TENSOR_SLICE_COPY_HPP
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "threadwise_tensor_slice_copy.hpp"
namespace ck {
template <index_t BlockSize,
class Float,
class SrcDesc,
class DstDesc,
class SrcLengths,
class SrcSubLengths,
class SrcClusterLengths,
class MapDst2Src,
class MapThreadCluster2SrcCluster,
index_t SrcDataPerRead,
index_t DstDataPerWrite>
struct BlockwiseTensorSliceReorderCopy_v3
{
static constexpr index_t nDim = SrcLengths::GetSize();
index_t mThreadSrcOffset;
index_t mThreadDstOffset;
__device__
BlockwiseTensorSliceReorderCopy_v3(Array<index_t, nDim> src_block_data_multi_id_begin,
Array<index_t, nDim> dst_block_data_multi_id_begin)
{
constexpr auto src_desc = SrcDesc{};
constexpr auto dst_desc = DstDesc{};
constexpr auto src_lengths = SrcLengths{};
constexpr auto map_dst2src = MapDst2Src{};
constexpr auto src_sub_lengths = SrcSubLengths{};
constexpr auto dst_sub_lengths = src_sub_lengths.ReorderGivenNew2Old(map_dst2src);
constexpr auto map_thread_cluster_2_src_cluster = MapThreadCluster2SrcCluster{};
constexpr auto src_cluster_lengths = SrcClusterLengths{};
constexpr auto thread_cluster_lengths =
src_cluster_lengths.ReorderGivenNew2Old(map_thread_cluster_2_src_cluster);
constexpr auto thread_cluster_desc =
make_ConstantTensorDescriptor_packed(thread_cluster_lengths);
// sanity check: data type
static_assert(is_same<Float, float>{}, "wrong! only support float for now!\n");
// sanity check: nDim
static_assert(SrcDesc::GetNumOfDimension() == nDim &&
DstDesc::GetNumOfDimension() == nDim && SrcLengths::GetSize() == nDim &&
SrcSubLengths::GetSize() == nDim &&
SrcClusterLengths::GetSize() == nDim && MapDst2Src::GetSize() == nDim &&
MapThreadCluster2SrcCluster::GetSize() == nDim,
"wrong! nDim is not consistent\n");
// sanity check: BlockSize
constexpr index_t num_active_thread = thread_cluster_desc.GetElementSize();
static_assert(BlockSize >= num_active_thread,
"wrong! BlockSize is not big enough for ThreadPerDims!");
// sanity check: work division
static_for<0, nDim, 1>{}([&](auto IDim) {
constexpr auto I = decltype(IDim){};
constexpr index_t src_len = src_lengths.Get(I);
constexpr index_t src_sub_len = src_sub_lengths.Get(I);
constexpr index_t src_cluster_len = src_cluster_lengths.Get(I);
static_assert(src_len % (src_sub_len * src_cluster_len) == 0,
"wrong! cannot evenly divide Src tensor lengths");
});
// sanity check: src read
static_assert(SrcDataPerRead == 1 || SrcDataPerRead == 2 || SrcDataPerRead == 4,
"wrong! only support SrcDataPerRead == 1, 2 or 4!\n");
static_assert(SrcDataPerRead == 1 || src_desc.GetStride(Number<nDim - 1>{}) == 1,
"wrong! only support src.stride(nDim-1) == 1 if SrcDataPerRead > 1!\n");
static_assert(src_sub_lengths.Get(Number<nDim - 1>{}) % SrcDataPerRead == 0,
"wrong! src_sub_lengths[nDim-1] % SrcDataPerRead != 0\n");
static_assert(src_desc.GetStride(Number<nDim - 2>{}) % SrcDataPerRead == 0,
"wrong! should satisfy src_desc.stride(nDim-2) % SrcDataPerRead == 0, to "
"keep alignment");
// sanity check: dst write
static_assert(DstDataPerWrite == 1 || DstDataPerWrite == 2 || DstDataPerWrite == 4,
"wrong! only support DstDataPerWrite == 1, 2 or 4!\n");
static_assert(DstDataPerWrite == 1 || dst_desc.GetStride(Number<nDim - 1>{}) == 1,
"wrong! only support dst.stride(nDim-1) == 1 if DstDataPerWrite > 1!\n");
static_assert(dst_sub_lengths.Get(Number<nDim - 1>{}) % DstDataPerWrite == 0,
"wrong! dst_sub_lengths[nDim-1] % DstDataPerWrite != 0\n");
static_assert(dst_desc.GetStride(Number<nDim - 2>{}) % DstDataPerWrite == 0,
"wrong! should satisfy dst_desc.stride(nDim-2) % DstDataPerWrite == 0, to "
"keep alignment");
// start dividing work
if(BlockSize > num_active_thread)
{
if(get_thread_local_1d_id() >= num_active_thread)
{
return;
}
}
const auto thread_multi_id =
thread_cluster_desc.GetMultiIndexFrom1dIndex(get_thread_local_1d_id());
// compiler: thread_multi_id, src_data_multi_id, dst_data_multi_id, will use separate
// regsiters, or only one copy???
auto src_data_multi_id =
reorder_array_given_old2new(thread_multi_id, map_thread_cluster_2_src_cluster);
static_for<0, nDim, 1>{}([&](auto IDim) {
constexpr index_t idim = IDim;
// compiler: will it really compute index here, or be merged with
// GetOffsetFromMultiIndex and
// optimized away???
src_data_multi_id(idim) *= src_sub_lengths.Get(IDim);
});
// compiler: will it really compute index here, or be merged with GetOffsetFromMultiIndex
// and
// optimized away???
const auto dst_data_multi_id = reorder_array_given_new2old(src_data_multi_id, map_dst2src);
mThreadSrcOffset =
src_desc.GetOffsetFromMultiIndex(src_data_multi_id + src_block_data_multi_id_begin);
mThreadDstOffset =
dst_desc.GetOffsetFromMultiIndex(dst_data_multi_id + dst_block_data_multi_id_begin);
#if 0
if(get_block_1d_id() == 0 && get_thread_local_1d_id() == 0)
{
print_ConstantTensorDescriptor(thread_cluster_desc, "thread_cluster_desc: ");
}
if(get_block_1d_id() == 0)
{
printf("id %5u %5u: "
"thread_multi_id: %u %u, "
"src_block_data_multi_id_begin: %u %u, "
"src_data_multi_id: %u %u, "
"mThreadSrcOffset %u, mThreadDstOffset %u \n",
get_block_1d_id(),
get_thread_local_1d_id(),
thread_multi_id[0],
thread_multi_id[1],
src_block_data_multi_id_begin[0],
src_block_data_multi_id_begin[1],
src_data_multi_id[0],
src_data_multi_id[1],
mThreadSrcOffset,
mThreadDstOffset);
}
#endif
}
__device__ static constexpr index_t GetRegisterClipboardSize()
{
constexpr auto thread_sub_tensor_lengths = SrcSubLengths{};
constexpr auto src_data_per_cluster_per_dims =
thread_sub_tensor_lengths * SrcClusterLengths{};
constexpr auto repeat_lengths = transform_sequences(
math::integer_divide_ceiler<index_t>{}, SrcLengths{}, src_data_per_cluster_per_dims);
constexpr auto thread_tensor_lengths = thread_sub_tensor_lengths * repeat_lengths;
constexpr auto thread_tensor_desc =
make_ConstantTensorDescriptor_packed(thread_tensor_lengths);
return thread_tensor_desc.GetElementSpace();
}
__device__ void RunLoadRegisterClipboard(const Float* __restrict__ p_src,
Float* __restrict__ p_clipboard) const
{
constexpr auto thread_sub_tensor_lengths = SrcSubLengths{};
constexpr auto src_data_per_cluster_per_dims =
thread_sub_tensor_lengths * SrcClusterLengths{};
constexpr auto repeat_lengths = transform_sequences(
math::integer_divide_ceiler<index_t>{}, SrcLengths{}, src_data_per_cluster_per_dims);
constexpr auto thread_tensor_lengths = thread_sub_tensor_lengths * repeat_lengths;
constexpr auto thread_tensor_desc =
make_ConstantTensorDescriptor_packed(thread_tensor_lengths);
static_ford<decltype(repeat_lengths)>{}([&](auto repeat_multi_id_) {
constexpr auto repeat_multi_id = decltype(repeat_multi_id_){};
constexpr auto src_data_multi_id = repeat_multi_id * src_data_per_cluster_per_dims;
constexpr auto clipboard_data_multi_id = repeat_multi_id * thread_sub_tensor_lengths;
constexpr index_t src_offset = SrcDesc{}.GetOffsetFromMultiIndex(src_data_multi_id);
constexpr index_t clipboard_offset =
thread_tensor_desc.GetOffsetFromMultiIndex(clipboard_data_multi_id);
threadwise_tensor_slice_copy(SrcDesc{},
p_src + src_offset + mThreadSrcOffset,
thread_tensor_desc,
p_clipboard + clipboard_offset,
thread_sub_tensor_lengths,
Number<SrcDataPerRead>{});
});
}
__device__ void RunStoreRegisterClipboard(const Float* __restrict__ p_clipboard,
Float* __restrict__ p_dst) const
{
constexpr auto thread_sub_tensor_lengths = SrcSubLengths{};
constexpr auto src_data_per_cluster_per_dims =
thread_sub_tensor_lengths * SrcClusterLengths{};
constexpr auto repeat_lengths = transform_sequences(
math::integer_divide_ceiler<index_t>{}, SrcLengths{}, src_data_per_cluster_per_dims);
constexpr auto thread_tensor_lengths = thread_sub_tensor_lengths * repeat_lengths;
constexpr auto thread_tensor_desc =
make_ConstantTensorDescriptor_packed(thread_tensor_lengths);
static_ford<decltype(repeat_lengths)>{}([&](auto repeat_multi_id_) {
constexpr auto repeat_multi_id = decltype(repeat_multi_id_){};
constexpr auto clipboard_data_multi_id = repeat_multi_id * thread_sub_tensor_lengths;
constexpr auto src_data_multi_id = repeat_multi_id * src_data_per_cluster_per_dims;
// reorder src_data_multi_id to get dst_data_multi_id
constexpr auto dst_data_multi_id = src_data_multi_id.ReorderGivenNew2Old(MapDst2Src{});
constexpr index_t clipboard_offset =
thread_tensor_desc.GetOffsetFromMultiIndex(clipboard_data_multi_id);
constexpr index_t dst_offset = DstDesc{}.GetOffsetFromMultiIndex(dst_data_multi_id);
// write in the order of dst
#if 1
threadwise_tensor_slice_copy_reorder_given_dst2src_v2(thread_tensor_desc,
p_clipboard + clipboard_offset,
DstDesc{},
p_dst + dst_offset +
mThreadDstOffset,
thread_sub_tensor_lengths,
MapDst2Src{});
#else
threadwise_tensor_slice_copy_reorder_given_dst2src_v3(thread_tensor_desc,
p_clipboard + clipboard_offset,
DstDesc{},
p_dst + dst_offset +
mThreadDstOffset,
thread_sub_tensor_lengths,
MapDst2Src{},
Number<DstDataPerWrite>{});
#endif
});
}
__device__ void Run(const Float* __restrict__ p_src, Float* __restrict__ p_dst) const
{
Float p_clipboard[GetRegisterClipboardSize()];
RunLoadRegisterClipboard(p_src, p_clipboard);
RunStoreRegisterClipboard(p_clipboard, p_dst);
}
// this function doesn't do santiy check on whether the slicing window is out of the boundary
// of the tensor being sliced
template <index_t IDim_, index_t StepSize, bool PositiveDirection>
__device__ void MoveSlicingWindowOnSourceTensor(
Number<IDim_>, Number<StepSize>, integral_constant<bool, PositiveDirection> direction)
{
constexpr auto IDim = Number<IDim_>{};
static_if<PositiveDirection>{}([&](auto fwd) {
mThreadSrcOffset += StepSize * fwd(SrcDesc{}).GetStride(IDim);
}).Else([&](auto fwd) { mThreadSrcOffset -= StepSize * fwd(SrcDesc{}).GetStride(IDim); });
}
};
} // namespace ck
#endif
composable_kernel/include/tensor_operation/threadwise_4d_tensor_op.hpp deleted 100644 → 0
View file @ f0716f5b
#ifndef CK_THREADWISE_4D_TENSOR_OP_HPP
#define CK_THREADWISE_4D_TENSOR_OP_HPP
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
namespace ck {
template <class Float, class Desc, class IDim, class NShift>
__device__ void threadwise_4d_tensor_shift_down(Desc, Float* __restrict__ p, IDim, NShift)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto desc = Desc{};
#if 0
if(get_thread_local_1d_id() == 0)
{
print_ConstantTensorDescriptor(desc, "threadwise_4d_tensor_shift_down: ");
}
#endif
constexpr index_t nshift = NShift::mValue;
constexpr index_t did0_end =
is_same<decltype(I0), IDim>{} ? desc.GetLength(I0) - nshift : desc.GetLength(I0);
constexpr index_t did1_end =
is_same<decltype(I1), IDim>{} ? desc.GetLength(I1) - nshift : desc.GetLength(I1);
constexpr index_t did2_end =
is_same<decltype(I2), IDim>{} ? desc.GetLength(I2) - nshift : desc.GetLength(I2);
constexpr index_t did3_end =
is_same<decltype(I3), IDim>{} ? desc.GetLength(I3) - nshift : desc.GetLength(I3);
for(index_t did0 = 0; did0 < did0_end; ++did0)
{
for(index_t did1 = 0; did1 < did1_end; ++did1)
{
for(index_t did2 = 0; did2 < did2_end; ++did2)
{
for(index_t did3 = 0; did3 < did3_end; ++did3)
{
const index_t dindex = desc.GetOffsetFromMultiIndex(did0, did1, did2, did3);
const index_t sindex = dindex + nshift * desc.GetStride(IDim{});
p[dindex] = p[sindex];
}
}
}
}
}
} // namespace ck
#endif
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment