mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-11 08:50:17 +00:00
refactor
This commit is contained in:
Chao Liu
@@ -1,7 +1,8 @@
|
||||
#pragma once
|
||||
#include <unistd.h>
|
||||
#include "device.hpp"
|
||||
#include "gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn.hip.hpp"
|
||||
#include "gridwise_convolution_wrapper.hip.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hip.hpp"
|
||||
|
||||
template <class T, class InDesc, class WeiDesc, class OutDesc>
|
||||
void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
|
||||
@@ -260,14 +261,6 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
|
||||
constexpr index_t HoPerThread = 1;
|
||||
constexpr index_t WoPerThread = 1;
|
||||
|
||||
constexpr index_t InBlockCopy_ThreadPerDimC = 8;
|
||||
constexpr index_t InBlockCopy_ThreadPerDimH = 2;
|
||||
constexpr index_t InBlockCopy_ThreadPerDimW = 2;
|
||||
constexpr index_t InBlockCopy_ThreadPerDimN = 4;
|
||||
constexpr index_t InBlockCopyDataPerRead = 4;
|
||||
|
||||
constexpr index_t WeiBlockCopyDataPerRead = 4;
|
||||
|
||||
constexpr index_t GemmMPerThreadSubC = 4;
|
||||
constexpr index_t GemmNPerThreadSubC = 4;
|
||||
constexpr index_t GemmMLevel0Cluster = 4;
|
||||
@@ -276,6 +269,13 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
|
||||
constexpr index_t GemmNLevel1Cluster = 4;
|
||||
constexpr index_t GemmKPerThreadLoop = 1;
|
||||
|
||||
constexpr index_t InBlockCopy_ThreadPerDimC = 8;
|
||||
constexpr index_t InBlockCopy_ThreadPerDimH = 2;
|
||||
constexpr index_t InBlockCopy_ThreadPerDimW = 2;
|
||||
constexpr index_t InBlockCopy_ThreadPerDimN = 4;
|
||||
constexpr index_t InBlockCopyDataPerRead = 4;
|
||||
|
||||
constexpr index_t WeiBlockCopyDataPerRead = 4;
|
||||
constexpr index_t OutThreadCopyDataPerWrite = 2;
|
||||
|
||||
constexpr index_t BlockSize = 128;
|
||||
@@ -289,43 +289,49 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
|
||||
|
||||
for(index_t i = 0; i < nrepeat; ++i)
|
||||
{
|
||||
float time = launch_kernel(
|
||||
gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn<GridSize,
|
||||
BlockSize,
|
||||
T,
|
||||
decltype(in_chwn_desc),
|
||||
decltype(wei_cyxk_desc),
|
||||
decltype(out_khwn_desc),
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
CPerBlock,
|
||||
HoPerBlock,
|
||||
WoPerBlock,
|
||||
NPerThread,
|
||||
KPerThread,
|
||||
HoPerThread,
|
||||
WoPerThread,
|
||||
Sequence<InBlockCopy_ThreadPerDimC,
|
||||
InBlockCopy_ThreadPerDimH,
|
||||
InBlockCopy_ThreadPerDimW,
|
||||
InBlockCopy_ThreadPerDimN>,
|
||||
InBlockCopyDataPerRead,
|
||||
WeiBlockCopyDataPerRead,
|
||||
GemmMPerThreadSubC,
|
||||
GemmNPerThreadSubC,
|
||||
GemmMLevel0Cluster,
|
||||
GemmNLevel0Cluster,
|
||||
GemmMLevel1Cluster,
|
||||
GemmNLevel1Cluster,
|
||||
GemmKPerThreadLoop,
|
||||
OutThreadCopyDataPerWrite>,
|
||||
dim3(GridSize),
|
||||
dim3(BlockSize),
|
||||
static_cast<T*>(in_chwn_device_buf.GetDeviceBuffer()),
|
||||
static_cast<T*>(wei_cyxk_device_buf.GetDeviceBuffer()),
|
||||
static_cast<T*>(out_khwn_device_buf.GetDeviceBuffer()));
|
||||
constexpr auto gridwise_conv =
|
||||
GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn<GridSize,
|
||||
BlockSize,
|
||||
T,
|
||||
decltype(in_chwn_desc),
|
||||
decltype(wei_cyxk_desc),
|
||||
decltype(out_khwn_desc),
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
CPerBlock,
|
||||
HoPerBlock,
|
||||
WoPerBlock,
|
||||
NPerThread,
|
||||
KPerThread,
|
||||
HoPerThread,
|
||||
WoPerThread,
|
||||
GemmMPerThreadSubC,
|
||||
GemmNPerThreadSubC,
|
||||
GemmMLevel0Cluster,
|
||||
GemmNLevel0Cluster,
|
||||
GemmMLevel1Cluster,
|
||||
GemmNLevel1Cluster,
|
||||
GemmKPerThreadLoop,
|
||||
Sequence<InBlockCopy_ThreadPerDimC,
|
||||
InBlockCopy_ThreadPerDimH,
|
||||
InBlockCopy_ThreadPerDimW,
|
||||
InBlockCopy_ThreadPerDimN>,
|
||||
InBlockCopyDataPerRead,
|
||||
WeiBlockCopyDataPerRead,
|
||||
OutThreadCopyDataPerWrite>{};
|
||||
|
||||
printf("Elapsed time : %f ms\n", time);
|
||||
float time = launch_kernel(run_gridwise_convolution<decltype(gridwise_conv), T>,
|
||||
dim3(GridSize),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
static_cast<T*>(in_chwn_device_buf.GetDeviceBuffer()),
|
||||
static_cast<T*>(wei_cyxk_device_buf.GetDeviceBuffer()),
|
||||
static_cast<T*>(out_khwn_device_buf.GetDeviceBuffer()));
|
||||
|
||||
printf("Elapsed time : %f ms, %f TFlop/s\n",
|
||||
time,
|
||||
(float)calculate_convolution_flops(InDesc{}, WeiDesc{}, OutDesc{}) /
|
||||
(std::size_t(1024) * 1024 * 1024 * 1024) / (time / 1000));
|
||||
usleep(std::min(time * 1000, float(10000)));
|
||||
}
|
||||
|
||||
|
||||
@@ -661,9 +661,9 @@ int main(int argc, char* argv[])
|
||||
device_direct_convolution_2_nchw_kcyx_nkhw
|
||||
#elif 0
|
||||
device_direct_convolution_2_vectorized_nchw_kcyx_nkhw
|
||||
#elif 0
|
||||
device_implicit_gemm_convolution_1_chwn_cyxk_khwn
|
||||
#elif 1
|
||||
device_implicit_gemm_convolution_1_chwn_cyxk_khwn
|
||||
#elif 0
|
||||
device_implicit_gemm_convolution_2_chwn_cyxk_khwn
|
||||
#endif
|
||||
(in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
|
||||
|
||||
@@ -164,11 +164,10 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
|
||||
n_repeat * NPerLevel1Cluster + n_in_sub_c};
|
||||
}
|
||||
|
||||
template <class FloatA, class FloatB, class FloatC, class Accumulator>
|
||||
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,
|
||||
Accumulator f_accum) const
|
||||
FloatC* __restrict__ p_c_thread) const
|
||||
{
|
||||
constexpr auto True = integral_constant<bool, true>{};
|
||||
constexpr auto False = integral_constant<bool, false>{};
|
||||
@@ -250,8 +249,7 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
|
||||
p_b_thread,
|
||||
c_thread_mtx,
|
||||
False,
|
||||
p_c_thread + ib * ThreadMatrixStrideC,
|
||||
f_accum);
|
||||
p_c_thread + ib * ThreadMatrixStrideC);
|
||||
|
||||
// read next batch of a, b
|
||||
if(BlockMatrixStrideA != 0)
|
||||
@@ -296,8 +294,7 @@ struct BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2
|
||||
p_b_thread,
|
||||
c_thread_mtx,
|
||||
False,
|
||||
p_c_thread + (BatchPerThread - 1) * ThreadMatrixStrideC,
|
||||
f_accum);
|
||||
p_c_thread + (BatchPerThread - 1) * ThreadMatrixStrideC);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -0,0 +1,328 @@
|
||||
#pragma once
|
||||
#include "common.hip.hpp"
|
||||
#include "ConstantTensorDescriptor.hip.hpp"
|
||||
#include "ConstantMatrixDescriptor.hip.hpp"
|
||||
#include "blockwise_4d_tensor_op.hip.hpp"
|
||||
#include "blockwise_2d_tensor_op.hip.hpp"
|
||||
#include "threadwise_nd_tensor_op.hip.hpp"
|
||||
#include "threadwise_4d_tensor_op.hip.hpp"
|
||||
#include "blockwise_batched_gemm.hip.hpp"
|
||||
|
||||
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,
|
||||
class InBlockCopyThreadPerDims,
|
||||
index_t InBlockCopyDataPerRead,
|
||||
index_t WeiBlockCopyDataPerRead,
|
||||
index_t OutThreadCopyDataPerWrite>
|
||||
struct GridwiseConvolutionImplicitGemm_v1_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
|
||||
{
|
||||
// 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 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;
|
||||
|
||||
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_ek_global_desc = make_ConstantTensorDescriptor(Sequence<C * Y * X, K>{});
|
||||
|
||||
// tensor view of blockwise input and weight in LDS
|
||||
// be careful of alignment
|
||||
constexpr auto in_chwn_block_desc = make_ConstantTensorDescriptor_aligned(
|
||||
Sequence<CPerBlock, HiPerBlock, WiPerBlock, NPerBlock>{},
|
||||
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_khwn_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_chwn_global_desc),
|
||||
decltype(in_chwn_block_desc),
|
||||
decltype(in_chwn_block_desc.GetLengths()),
|
||||
InBlockCopyThreadPerDims,
|
||||
InBlockCopyDataPerRead>{};
|
||||
|
||||
// blockwise wei copy
|
||||
// format is [CPerBlock*Y*X,KPerBlock]
|
||||
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>{};
|
||||
|
||||
// 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_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>{},
|
||||
Number<out_khwn_thread_desc.GetStride(I1)>{});
|
||||
|
||||
const auto blockwise_batch_gemm =
|
||||
BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2<
|
||||
BlockSize,
|
||||
decltype(a_cxk_block_mtx_desc),
|
||||
decltype(b_cxwn_block_mtx_desc),
|
||||
decltype(c_kxwn_thread_mtx_desc),
|
||||
0,
|
||||
in_chwn_block_desc.GetStride(I1),
|
||||
out_khwn_thread_desc.GetStride(I1),
|
||||
HoPerBlock,
|
||||
GemmMPerThreadSubC,
|
||||
GemmNPerThreadSubC,
|
||||
GemmMLevel0Cluster,
|
||||
GemmNLevel0Cluster,
|
||||
GemmMLevel1Cluster,
|
||||
GemmNLevel1Cluster,
|
||||
GemmKPerThreadLoop,
|
||||
HoPerThread>{};
|
||||
|
||||
// LDS: be careful of alignment
|
||||
constexpr index_t in_block_element_size =
|
||||
in_chwn_block_desc.GetElementSpace(Number<InBlockCopyDataPerRead>{});
|
||||
|
||||
constexpr index_t wei_block_element_size =
|
||||
wei_cyxk_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)];
|
||||
|
||||
// register
|
||||
Float p_out_thread[out_khwn_thread_desc.GetElementSpace()];
|
||||
|
||||
// set threadwise output tensor to 0
|
||||
threadwise_4d_tensor_set_zero(out_khwn_thread_desc, p_out_thread);
|
||||
|
||||
const Float* p_in_global_block_begin =
|
||||
p_in_global +
|
||||
in_chwn_global_desc.Get1dIndex(
|
||||
0, hi_block_data_begin, wi_block_data_begin, n_block_data_begin);
|
||||
|
||||
const Float* p_wei_global_block_begin =
|
||||
p_wei_global + wei_cyxk_global_desc.Get1dIndex(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_begin += CPerBlock * in_chwn_global_desc.GetStride(I0),
|
||||
p_wei_global_block_begin += CPerBlock * wei_cyxk_global_desc.GetStride(I0),
|
||||
__syncthreads())
|
||||
{
|
||||
// input: global mem to LDS
|
||||
blockwise_in_copy.Run(p_in_global_block_begin, p_in_block);
|
||||
|
||||
// weight: global mem to LDS
|
||||
blockwise_wei_copy.Run(p_wei_global_block_begin, p_wei_block);
|
||||
|
||||
__syncthreads();
|
||||
|
||||
// a series of batched GEMM
|
||||
for(index_t y = 0; y < Y; ++y)
|
||||
{
|
||||
for(index_t x = 0; x < X; ++x)
|
||||
{
|
||||
blockwise_batch_gemm.Run(p_wei_block +
|
||||
wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
|
||||
p_in_block + in_chwn_block_desc.Get1dIndex(0, y, x, 0),
|
||||
p_out_thread);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// output: register to global mem,
|
||||
#if 0
|
||||
const auto c_thread_mtx_begin =
|
||||
blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
|
||||
|
||||
for(index_t k = 0; k < out_khwn_thread_desc.GetLength(I0); ++k)
|
||||
{
|
||||
for(index_t ho = 0; ho < out_khwn_thread_desc.GetLength(I1); ++ho)
|
||||
{
|
||||
for(index_t wo = 0; wo < out_khwn_thread_desc.GetLength(I2); ++wo)
|
||||
{
|
||||
for(index_t n = 0; n < out_khwn_thread_desc.GetLength(I3); ++n)
|
||||
{
|
||||
const index_t b = out_khwn_thread_desc.Get1dIndex(0, 0, wo, n);
|
||||
|
||||
const auto c_thread_mtx_distance =
|
||||
blockwise_batch_gemm.GetDistanceFromBeginOfThreadMatrixC(ho, k, b);
|
||||
|
||||
const index_t ho_thread =
|
||||
c_thread_mtx_begin.batch + c_thread_mtx_distance.batch;
|
||||
const index_t k_thread = c_thread_mtx_begin.row + c_thread_mtx_distance.row;
|
||||
const index_t b_thread = c_thread_mtx_begin.col + c_thread_mtx_distance.col;
|
||||
|
||||
const index_t wo_thread = b_thread / NPerBlock;
|
||||
const index_t n_thread = b_thread % NPerBlock;
|
||||
|
||||
p_out_global[out_khwn_global_desc.Get1dIndex(k_block_data_begin + k_thread,
|
||||
ho_block_data_begin + ho_thread,
|
||||
wo_block_data_begin + wo_thread,
|
||||
n_block_data_begin + n_thread)] =
|
||||
p_out_thread[out_khwn_thread_desc.Get1dIndex(k, ho, wo, n)];
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
#elif 1
|
||||
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 * wo_thread_data_begin;
|
||||
|
||||
// this is for v2 GEMM
|
||||
// output is a 8d tensor
|
||||
if(NPerThread < NPerBlock && WoPerThread == 1)
|
||||
{
|
||||
constexpr index_t N1_ = GemmNPerThreadSubC;
|
||||
constexpr index_t W1_ = WoPerBlock / ((WoPerThread * NPerThread) / GemmNPerThreadSubC);
|
||||
constexpr index_t K2_ = GemmMPerThreadSubC;
|
||||
constexpr index_t K1_ = KPerBlock / KPerThread;
|
||||
|
||||
constexpr auto out_8d_global_desc = make_ConstantTensorDescriptor(
|
||||
Sequence<K / (K1_ * K2_), K1_, K2_, Ho, Wo / W1_, W1_, N / N1_, N1_>{});
|
||||
|
||||
constexpr auto out_8d_thread_desc =
|
||||
make_ConstantTensorDescriptor(Sequence<KPerBlock / (K1_ * K2_),
|
||||
1,
|
||||
K2_,
|
||||
HoPerThread,
|
||||
WoPerBlock / W1_,
|
||||
1,
|
||||
1,
|
||||
N1_>{});
|
||||
|
||||
#if 0
|
||||
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
|
||||
{
|
||||
print_ConstantTensorDescriptor(out_khwn_thread_desc, "out_khwn_thread_desc");
|
||||
print_ConstantTensorDescriptor(out_8d_thread_desc, "out_8d_thread_desc");
|
||||
|
||||
print_ConstantTensorDescriptor(out_khwn_global_desc, "out_khwn_global_desc");
|
||||
print_ConstantTensorDescriptor(out_8d_global_desc, "out_8d_global_desc");
|
||||
}
|
||||
#endif
|
||||
|
||||
threadwise_8d_tensor_copy(
|
||||
out_8d_thread_desc,
|
||||
p_out_thread,
|
||||
out_8d_global_desc,
|
||||
p_out_global +
|
||||
out_khwn_global_desc.Get1dIndex(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_8d_thread_desc.GetLengths(),
|
||||
Number<OutThreadCopyDataPerWrite>{});
|
||||
}
|
||||
else if(NPerThread == NPerBlock)
|
||||
{
|
||||
// not implemented yet
|
||||
assert(false);
|
||||
}
|
||||
else
|
||||
{
|
||||
assert(false);
|
||||
}
|
||||
#endif
|
||||
}
|
||||
};
|
||||
@@ -1,318 +0,0 @@
|
||||
#pragma once
|
||||
#include "common.hip.hpp"
|
||||
#include "ConstantTensorDescriptor.hip.hpp"
|
||||
#include "ConstantMatrixDescriptor.hip.hpp"
|
||||
#include "blockwise_4d_tensor_op.hip.hpp"
|
||||
#include "blockwise_2d_tensor_op.hip.hpp"
|
||||
#include "threadwise_nd_tensor_op.hip.hpp"
|
||||
#include "threadwise_4d_tensor_op.hip.hpp"
|
||||
#include "blockwise_batched_gemm.hip.hpp"
|
||||
|
||||
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,
|
||||
class InBlockCopyThreadPerDims,
|
||||
index_t InBlockCopyDataPerRead,
|
||||
index_t WeiBlockCopyDataPerRead,
|
||||
index_t GemmMPerThreadSubC,
|
||||
index_t GemmNPerThreadSubC,
|
||||
index_t GemmMLevel0Cluster,
|
||||
index_t GemmNLevel0Cluster,
|
||||
index_t GemmMLevel1Cluster,
|
||||
index_t GemmNLevel1Cluster,
|
||||
index_t GemmKPerThreadLoop,
|
||||
index_t OutThreadCopyDataPerWrite>
|
||||
__global__ void
|
||||
gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn(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 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;
|
||||
|
||||
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_ek_global_desc = make_ConstantTensorDescriptor(Sequence<C * Y * X, K>{});
|
||||
|
||||
// tensor view of blockwise input and weight in LDS
|
||||
// be careful of alignment
|
||||
constexpr auto in_chwn_block_desc = make_ConstantTensorDescriptor_aligned(
|
||||
Sequence<CPerBlock, HiPerBlock, WiPerBlock, NPerBlock>{}, 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_khwn_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_chwn_global_desc),
|
||||
decltype(in_chwn_block_desc),
|
||||
decltype(in_chwn_block_desc.GetLengths()),
|
||||
InBlockCopyThreadPerDims,
|
||||
InBlockCopyDataPerRead>{};
|
||||
|
||||
// blockwise wei copy
|
||||
// format is [CPerBlock*Y*X,KPerBlock]
|
||||
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>{};
|
||||
|
||||
// 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_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>{},
|
||||
Number<out_khwn_thread_desc.GetStride(I1)>{});
|
||||
|
||||
const auto blockwise_batch_gemm = BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2<
|
||||
BlockSize,
|
||||
decltype(a_cxk_block_mtx_desc),
|
||||
decltype(b_cxwn_block_mtx_desc),
|
||||
decltype(c_kxwn_thread_mtx_desc),
|
||||
0,
|
||||
in_chwn_block_desc.GetStride(I1),
|
||||
out_khwn_thread_desc.GetStride(I1),
|
||||
HoPerBlock,
|
||||
GemmMPerThreadSubC,
|
||||
GemmNPerThreadSubC,
|
||||
GemmMLevel0Cluster,
|
||||
GemmNLevel0Cluster,
|
||||
GemmMLevel1Cluster,
|
||||
GemmNLevel1Cluster,
|
||||
GemmKPerThreadLoop,
|
||||
HoPerThread>{};
|
||||
|
||||
// LDS: be careful of alignment
|
||||
constexpr index_t in_block_element_size =
|
||||
in_chwn_block_desc.GetElementSpace(Number<InBlockCopyDataPerRead>{});
|
||||
|
||||
constexpr index_t wei_block_element_size =
|
||||
wei_cyxk_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)];
|
||||
|
||||
// register
|
||||
Float p_out_thread[out_khwn_thread_desc.GetElementSpace()];
|
||||
|
||||
// set threadwise output tensor to 0
|
||||
threadwise_4d_tensor_set_zero(out_khwn_thread_desc, p_out_thread);
|
||||
|
||||
const Float* p_in_global_block_begin =
|
||||
p_in_global +
|
||||
in_chwn_global_desc.Get1dIndex(
|
||||
0, hi_block_data_begin, wi_block_data_begin, n_block_data_begin);
|
||||
|
||||
const Float* p_wei_global_block_begin =
|
||||
p_wei_global + wei_cyxk_global_desc.Get1dIndex(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_begin += CPerBlock * in_chwn_global_desc.GetStride(I0),
|
||||
p_wei_global_block_begin += CPerBlock * wei_cyxk_global_desc.GetStride(I0),
|
||||
__syncthreads())
|
||||
{
|
||||
// input: global mem to LDS
|
||||
blockwise_in_copy.Run(p_in_global_block_begin, p_in_block);
|
||||
|
||||
// weight: global mem to LDS
|
||||
blockwise_wei_copy.Run(p_wei_global_block_begin, p_wei_block);
|
||||
|
||||
__syncthreads();
|
||||
|
||||
// a series of batched GEMM
|
||||
for(index_t y = 0; y < Y; ++y)
|
||||
{
|
||||
for(index_t x = 0; x < X; ++x)
|
||||
{
|
||||
#if 0
|
||||
blockwise_batch_gemm.Run
|
||||
#elif 1
|
||||
blockwise_batch_gemm.Run_v3
|
||||
#endif
|
||||
(p_wei_block + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
|
||||
p_in_block + in_chwn_block_desc.Get1dIndex(0, y, x, 0),
|
||||
p_out_thread,
|
||||
[](auto& acc, const auto&& v) { acc += v; });
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// output: register to global mem,
|
||||
#if 0
|
||||
const auto c_thread_mtx_begin =
|
||||
blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
|
||||
|
||||
for(index_t k = 0; k < out_khwn_thread_desc.GetLength(I0); ++k)
|
||||
{
|
||||
for(index_t ho = 0; ho < out_khwn_thread_desc.GetLength(I1); ++ho)
|
||||
{
|
||||
for(index_t wo = 0; wo < out_khwn_thread_desc.GetLength(I2); ++wo)
|
||||
{
|
||||
for(index_t n = 0; n < out_khwn_thread_desc.GetLength(I3); ++n)
|
||||
{
|
||||
const index_t b = out_khwn_thread_desc.Get1dIndex(0, 0, wo, n);
|
||||
|
||||
const auto c_thread_mtx_distance =
|
||||
blockwise_batch_gemm.GetDistanceFromBeginOfThreadMatrixC(ho, k, b);
|
||||
|
||||
const index_t ho_thread =
|
||||
c_thread_mtx_begin.batch + c_thread_mtx_distance.batch;
|
||||
const index_t k_thread = c_thread_mtx_begin.row + c_thread_mtx_distance.row;
|
||||
const index_t b_thread = c_thread_mtx_begin.col + c_thread_mtx_distance.col;
|
||||
|
||||
const index_t wo_thread = b_thread / NPerBlock;
|
||||
const index_t n_thread = b_thread % NPerBlock;
|
||||
|
||||
p_out_global[out_khwn_global_desc.Get1dIndex(k_block_data_begin + k_thread,
|
||||
ho_block_data_begin + ho_thread,
|
||||
wo_block_data_begin + wo_thread,
|
||||
n_block_data_begin + n_thread)] =
|
||||
p_out_thread[out_khwn_thread_desc.Get1dIndex(k, ho, wo, n)];
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
#elif 1
|
||||
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 * wo_thread_data_begin;
|
||||
|
||||
// this is for v2 GEMM
|
||||
// output is a 8d tensor
|
||||
if(NPerThread < NPerBlock && WoPerThread == 1)
|
||||
{
|
||||
constexpr index_t N1_ = GemmNPerThreadSubC;
|
||||
constexpr index_t W1_ = WoPerBlock / ((WoPerThread * NPerThread) / GemmNPerThreadSubC);
|
||||
constexpr index_t K2_ = GemmMPerThreadSubC;
|
||||
constexpr index_t K1_ = KPerBlock / KPerThread;
|
||||
|
||||
constexpr auto out_8d_global_desc = make_ConstantTensorDescriptor(
|
||||
Sequence<K / (K1_ * K2_), K1_, K2_, Ho, Wo / W1_, W1_, N / N1_, N1_>{});
|
||||
|
||||
constexpr auto out_8d_thread_desc = make_ConstantTensorDescriptor(
|
||||
Sequence<KPerBlock / (K1_ * K2_), 1, K2_, HoPerThread, WoPerBlock / W1_, 1, 1, N1_>{});
|
||||
|
||||
#if 0
|
||||
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
|
||||
{
|
||||
print_ConstantTensorDescriptor(out_khwn_thread_desc, "out_khwn_thread_desc");
|
||||
print_ConstantTensorDescriptor(out_8d_thread_desc, "out_8d_thread_desc");
|
||||
|
||||
print_ConstantTensorDescriptor(out_khwn_global_desc, "out_khwn_global_desc");
|
||||
print_ConstantTensorDescriptor(out_8d_global_desc, "out_8d_global_desc");
|
||||
}
|
||||
#endif
|
||||
|
||||
threadwise_8d_tensor_copy(
|
||||
out_8d_thread_desc,
|
||||
p_out_thread,
|
||||
out_8d_global_desc,
|
||||
p_out_global +
|
||||
out_khwn_global_desc.Get1dIndex(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_8d_thread_desc.GetLengths(),
|
||||
Number<OutThreadCopyDataPerWrite>{});
|
||||
}
|
||||
else if(NPerThread == NPerBlock)
|
||||
{
|
||||
// not implemented yet
|
||||
assert(false);
|
||||
}
|
||||
else
|
||||
{
|
||||
assert(false);
|
||||
}
|
||||
#endif
|
||||
}
|
||||
Reference in New Issue
Block a user