mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-03 21:21:22 +00:00
Update Group convolution (#341)
* add conv oddC * update example * update example * fix bug in example * fix bug in group conv example
This commit is contained in:
Chao Liu
@@ -1,876 +0,0 @@
|
||||
#pragma once
|
||||
|
||||
#include <iostream>
|
||||
#include <sstream>
|
||||
|
||||
#include "ck/utility/common_header.hpp"
|
||||
#include "ck/tensor_description/tensor_descriptor.hpp"
|
||||
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_batched_gemm_c_permute.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
|
||||
#include "ck/host_utility/device_prop.hpp"
|
||||
#include "ck/host_utility/kernel_launch.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
|
||||
/*
|
||||
* \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
|
||||
*
|
||||
* \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
|
||||
* given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
|
||||
* strided batched, but we can easily extend to other layouts. The returned offset can be either \p
|
||||
* index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
|
||||
* limitations.
|
||||
*
|
||||
* \tparam Block2CTileMap Block2CTileMap::CalculateBottomIndex() takes in id of a workgroup and
|
||||
* returns the 2D index of the tile that it computes. \see
|
||||
* GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
|
||||
*
|
||||
* \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
|
||||
* tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
|
||||
* descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
|
||||
* device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
|
||||
* DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
|
||||
* pointer offset into \p ComputePtrOffsetOfStridedBatch.
|
||||
*
|
||||
* \note \p Block2CTileMap allows customized mapping between a workgroup and the C-tile it computes.
|
||||
* Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
|
||||
* realize BatchedGemmCPermute and GroupedGemm (and the corresponding GEMM fusion).
|
||||
*
|
||||
*/
|
||||
template <typename GridwiseGemm,
|
||||
typename FloatAB,
|
||||
typename FloatC,
|
||||
typename AGridDesc_AK0_M_AK1,
|
||||
typename BGridDesc_BK0_N_BK1,
|
||||
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CDEElementwiseOperation,
|
||||
typename ComputePtrOffsetOfBatch,
|
||||
typename Block2CTileMap,
|
||||
bool HasMainKBlockLoop>
|
||||
__global__ void
|
||||
#if CK_USE_LAUNCH_BOUNDS
|
||||
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
|
||||
#endif
|
||||
kernel_batched_gemm_c_permute_xdl(const FloatAB* __restrict__ p_a_grid,
|
||||
const FloatAB* __restrict__ p_b_grid,
|
||||
FloatC* __restrict__ p_e_grid,
|
||||
const index_t batch_count,
|
||||
const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
|
||||
const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
|
||||
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
|
||||
c_grid_desc_mblock_mperblock_nblock_nperblock,
|
||||
const AElementwiseOperation a_element_op,
|
||||
const BElementwiseOperation b_element_op,
|
||||
const CDEElementwiseOperation cde_element_op,
|
||||
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
|
||||
const Block2CTileMap block_2_ctile_map)
|
||||
{
|
||||
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
|
||||
const index_t num_blocks_per_batch =
|
||||
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
|
||||
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
|
||||
|
||||
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
|
||||
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
|
||||
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
|
||||
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
|
||||
const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
|
||||
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
|
||||
|
||||
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
|
||||
|
||||
GridwiseGemm::template Run<HasMainKBlockLoop>(
|
||||
p_a_grid + a_batch_offset,
|
||||
p_b_grid + b_batch_offset,
|
||||
ck::Tuple<>{},
|
||||
p_e_grid + c_batch_offset,
|
||||
p_shared,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op,
|
||||
a_grid_desc_k0_m_k1,
|
||||
b_grid_desc_k0_n_k1,
|
||||
ck::StaticallyIndexedArray<
|
||||
typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
0>{},
|
||||
c_grid_desc_mblock_mperblock_nblock_nperblock,
|
||||
block_2_ctile_map);
|
||||
#else
|
||||
ignore = p_a_grid;
|
||||
ignore = p_b_grid;
|
||||
ignore = p_e_grid;
|
||||
ignore = batch_count;
|
||||
ignore = a_grid_desc_k0_m_k1;
|
||||
ignore = b_grid_desc_k0_n_k1;
|
||||
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
|
||||
ignore = a_element_op;
|
||||
ignore = b_element_op;
|
||||
ignore = cde_element_op;
|
||||
ignore = compute_ptr_offset_of_batch;
|
||||
ignore = block_2_ctile_map;
|
||||
#endif
|
||||
}
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename DELayout,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename GemmAccDataType,
|
||||
typename CShuffleDataType,
|
||||
typename DsDataType,
|
||||
typename EDataType,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CDEElementwiseOperation,
|
||||
GemmSpecialization GemmSpec,
|
||||
index_t NumGemmKPrefetchStage,
|
||||
index_t BlockSize,
|
||||
index_t MPerBlock,
|
||||
index_t NPerBlock,
|
||||
index_t KPerBlock,
|
||||
index_t AK1,
|
||||
index_t BK1,
|
||||
index_t MPerXDL,
|
||||
index_t NPerXDL,
|
||||
index_t MXdlPerWave,
|
||||
index_t NXdlPerWave,
|
||||
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
typename ABlockTransferThreadClusterArrangeOrder,
|
||||
typename ABlockTransferSrcAccessOrder,
|
||||
index_t ABlockTransferSrcVectorDim,
|
||||
index_t ABlockTransferSrcScalarPerVector,
|
||||
index_t ABlockTransferDstScalarPerVector_AK1,
|
||||
bool ABlockLdsExtraM,
|
||||
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
|
||||
typename BBlockTransferThreadClusterArrangeOrder,
|
||||
typename BBlockTransferSrcAccessOrder,
|
||||
index_t BBlockTransferSrcVectorDim,
|
||||
index_t BBlockTransferSrcScalarPerVector,
|
||||
index_t BBlockTransferDstScalarPerVector_BK1,
|
||||
bool BBlockLdsExtraN,
|
||||
index_t CShuffleMXdlPerWavePerShuffle,
|
||||
index_t CShuffleNXdlPerWavePerShuffle,
|
||||
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
index_t CDEBlockTransferScalarPerVector_NPerBlock,
|
||||
LoopScheduler LoopSched = make_default_loop_scheduler()>
|
||||
struct DeviceBatchedGemmCPermuteXdl : public DeviceBatchedGemmCPermute<ALayout,
|
||||
BLayout,
|
||||
DELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
EDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation>
|
||||
{
|
||||
|
||||
using DeviceOp = DeviceBatchedGemmCPermuteXdl;
|
||||
|
||||
static constexpr auto I0 = Number<0>{};
|
||||
static constexpr auto I1 = Number<1>{};
|
||||
static constexpr auto I2 = Number<2>{};
|
||||
|
||||
static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
|
||||
{
|
||||
const auto a_grid_desc_mraw_kraw = [&]() {
|
||||
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
|
||||
{
|
||||
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
|
||||
make_tuple(StrideA, I1));
|
||||
}
|
||||
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
|
||||
{
|
||||
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
|
||||
make_tuple(I1, StrideA));
|
||||
}
|
||||
}();
|
||||
|
||||
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
|
||||
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
|
||||
|
||||
const auto MPad = M - MRaw;
|
||||
const auto KPad = K - KRaw;
|
||||
|
||||
if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
|
||||
GemmSpec == GemmSpecialization::MNKPadding)
|
||||
{
|
||||
// pad both M and K
|
||||
assert(K % AK1 == 0);
|
||||
|
||||
const auto AK0 = K / AK1;
|
||||
|
||||
const auto a_grid_desc_m_k =
|
||||
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
|
||||
make_tuple(make_right_pad_transform(MRaw, MPad),
|
||||
make_right_pad_transform(KRaw, KPad)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}));
|
||||
|
||||
const auto a_grid_desc_ak0_m_ak1 =
|
||||
transform_tensor_descriptor(a_grid_desc_m_k,
|
||||
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
|
||||
make_pass_through_transform(M)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
return a_grid_desc_ak0_m_ak1;
|
||||
}
|
||||
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
|
||||
GemmSpec == GemmSpecialization::MNPadding)
|
||||
{
|
||||
// pad M, but not K
|
||||
assert(KRaw % AK1 == 0);
|
||||
|
||||
const auto AK0 = KRaw / AK1;
|
||||
|
||||
const auto a_grid_desc_ak0_m_ak1 =
|
||||
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
|
||||
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
|
||||
make_right_pad_transform(MRaw, MPad)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
return a_grid_desc_ak0_m_ak1;
|
||||
}
|
||||
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
|
||||
GemmSpec == GemmSpecialization::NKPadding)
|
||||
{
|
||||
// pad K, but not M
|
||||
assert(K % AK1 == 0);
|
||||
|
||||
const auto AK0 = K / AK1;
|
||||
|
||||
const auto a_grid_desc_m_k = transform_tensor_descriptor(
|
||||
a_grid_desc_mraw_kraw,
|
||||
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}));
|
||||
|
||||
const auto a_grid_desc_ak0_m_ak1 =
|
||||
transform_tensor_descriptor(a_grid_desc_m_k,
|
||||
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
|
||||
make_pass_through_transform(MRaw)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
return a_grid_desc_ak0_m_ak1;
|
||||
}
|
||||
else
|
||||
{
|
||||
// not pad M or K
|
||||
assert(KRaw % AK1 == 0);
|
||||
|
||||
const auto AK0 = KRaw / AK1;
|
||||
|
||||
const auto a_grid_desc_ak0_m_ak1 =
|
||||
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
|
||||
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
|
||||
make_pass_through_transform(MRaw)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
return a_grid_desc_ak0_m_ak1;
|
||||
}
|
||||
}
|
||||
|
||||
static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
|
||||
{
|
||||
const auto b_grid_desc_nraw_kraw = [&]() {
|
||||
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
|
||||
{
|
||||
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
|
||||
make_tuple(I1, StrideB));
|
||||
}
|
||||
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
|
||||
{
|
||||
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
|
||||
make_tuple(StrideB, I1));
|
||||
}
|
||||
}();
|
||||
|
||||
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
|
||||
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
|
||||
|
||||
const auto NPad = N - NRaw;
|
||||
const auto KPad = K - KRaw;
|
||||
|
||||
if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
|
||||
GemmSpec == GemmSpecialization::MNKPadding)
|
||||
{
|
||||
// pad both N and K
|
||||
assert(K % BK1 == 0);
|
||||
|
||||
const auto BK0 = K / BK1;
|
||||
|
||||
const auto b_grid_desc_n_k =
|
||||
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
|
||||
make_tuple(make_right_pad_transform(NRaw, NPad),
|
||||
make_right_pad_transform(KRaw, KPad)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}));
|
||||
|
||||
const auto b_grid_desc_bk0_n_bk1 =
|
||||
transform_tensor_descriptor(b_grid_desc_n_k,
|
||||
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
|
||||
make_pass_through_transform(N)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
return b_grid_desc_bk0_n_bk1;
|
||||
}
|
||||
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
|
||||
GemmSpec == GemmSpecialization::MNPadding)
|
||||
{
|
||||
// pad N, but not K
|
||||
assert(KRaw % BK1 == 0);
|
||||
|
||||
const auto BK0 = KRaw / BK1;
|
||||
|
||||
const auto b_grid_desc_bk0_n_bk1 =
|
||||
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
|
||||
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
|
||||
make_right_pad_transform(NRaw, NPad)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
return b_grid_desc_bk0_n_bk1;
|
||||
}
|
||||
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
|
||||
GemmSpec == GemmSpecialization::MKPadding)
|
||||
{
|
||||
// pad K, but not N
|
||||
assert(K % BK1 == 0);
|
||||
|
||||
const auto BK0 = K / BK1;
|
||||
|
||||
const auto b_grid_desc_n_k = transform_tensor_descriptor(
|
||||
b_grid_desc_nraw_kraw,
|
||||
make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}));
|
||||
|
||||
const auto b_grid_desc_bk0_n_bk1 =
|
||||
transform_tensor_descriptor(b_grid_desc_n_k,
|
||||
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
|
||||
make_pass_through_transform(NRaw)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
return b_grid_desc_bk0_n_bk1;
|
||||
}
|
||||
else
|
||||
{
|
||||
// not pad N or K
|
||||
assert(KRaw % BK1 == 0);
|
||||
|
||||
const auto BK0 = KRaw / BK1;
|
||||
|
||||
const auto b_grid_desc_bk0_n_bk1 =
|
||||
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
|
||||
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
|
||||
make_pass_through_transform(NRaw)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
return b_grid_desc_bk0_n_bk1;
|
||||
}
|
||||
}
|
||||
|
||||
static auto
|
||||
MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t stride_M, index_t stride_N)
|
||||
{
|
||||
const auto c_grid_desc_mraw_nraw = [&]() {
|
||||
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
|
||||
make_tuple(stride_M, stride_N));
|
||||
}();
|
||||
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
|
||||
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
|
||||
|
||||
const auto MPad = M - MRaw;
|
||||
const auto NPad = N - NRaw;
|
||||
|
||||
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
|
||||
GemmSpec == GemmSpecialization::MNKPadding)
|
||||
{
|
||||
// pad M and N
|
||||
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
|
||||
make_tuple(make_right_pad_transform(MRaw, MPad),
|
||||
make_right_pad_transform(NRaw, NPad)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}));
|
||||
}
|
||||
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
|
||||
GemmSpec == GemmSpecialization::MKPadding)
|
||||
{
|
||||
// pad M, but not N
|
||||
return transform_tensor_descriptor(
|
||||
c_grid_desc_mraw_nraw,
|
||||
make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}));
|
||||
}
|
||||
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
|
||||
GemmSpec == GemmSpecialization::NKPadding)
|
||||
{
|
||||
// pad N, but not M
|
||||
return transform_tensor_descriptor(
|
||||
c_grid_desc_mraw_nraw,
|
||||
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}));
|
||||
}
|
||||
else
|
||||
{
|
||||
// not pad M or N
|
||||
return c_grid_desc_mraw_nraw;
|
||||
}
|
||||
}
|
||||
|
||||
static auto MakeEGridDescriptor_G0_G1_M_N(index_t G0,
|
||||
index_t G1,
|
||||
index_t MRaw,
|
||||
index_t NRaw,
|
||||
index_t stride_G0,
|
||||
index_t stride_G1,
|
||||
index_t stride_M,
|
||||
index_t stride_N)
|
||||
{
|
||||
const auto e_grid_desc_g0_g1_mraw_nraw = [&]() {
|
||||
return make_naive_tensor_descriptor(
|
||||
make_tuple(G0, G1, MRaw, NRaw),
|
||||
make_tuple(stride_G0, stride_G1, stride_M, stride_N));
|
||||
}();
|
||||
|
||||
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
|
||||
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
|
||||
|
||||
const auto MPad = M - MRaw;
|
||||
const auto NPad = N - NRaw;
|
||||
|
||||
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
|
||||
GemmSpec == GemmSpecialization::MNKPadding)
|
||||
{
|
||||
// pad M and N
|
||||
return transform_tensor_descriptor(
|
||||
e_grid_desc_g0_g1_mraw_nraw,
|
||||
make_tuple(make_pass_through_transform(G0),
|
||||
make_pass_through_transform(G1),
|
||||
make_right_pad_transform(MRaw, MPad),
|
||||
make_right_pad_transform(NRaw, NPad)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
|
||||
}
|
||||
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
|
||||
GemmSpec == GemmSpecialization::MKPadding)
|
||||
{
|
||||
// pad M, but not N
|
||||
return transform_tensor_descriptor(
|
||||
e_grid_desc_g0_g1_mraw_nraw,
|
||||
make_tuple(make_pass_through_transform(G0),
|
||||
make_pass_through_transform(G1),
|
||||
make_right_pad_transform(MRaw, MPad),
|
||||
make_pass_through_transform(NRaw)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
|
||||
}
|
||||
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
|
||||
GemmSpec == GemmSpecialization::NKPadding)
|
||||
{
|
||||
// pad N, but not M
|
||||
return transform_tensor_descriptor(
|
||||
e_grid_desc_g0_g1_mraw_nraw,
|
||||
make_tuple(make_pass_through_transform(G0),
|
||||
make_pass_through_transform(G1),
|
||||
make_pass_through_transform(MRaw),
|
||||
make_right_pad_transform(NRaw, NPad)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
|
||||
}
|
||||
else
|
||||
{
|
||||
// not pad M or N
|
||||
return e_grid_desc_g0_g1_mraw_nraw;
|
||||
}
|
||||
}
|
||||
|
||||
using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
|
||||
using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
|
||||
using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N(1, 1, 1, 1));
|
||||
using EGridDesc_G0_G1_M_N = decltype(MakeEGridDescriptor_G0_G1_M_N(1, 1, 1, 1, 1, 1, 1, 1));
|
||||
|
||||
struct ComputePtrOffsetOfStridedBatch
|
||||
{
|
||||
ComputePtrOffsetOfStridedBatch(index_t Batchstride_A,
|
||||
index_t Batchstride_B,
|
||||
EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n)
|
||||
: Batchstride_A_(Batchstride_A),
|
||||
Batchstride_B_(Batchstride_B),
|
||||
e_grid_desc_g0_g1_m_n_(e_grid_desc_g0_g1_m_n)
|
||||
{
|
||||
}
|
||||
|
||||
__host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
|
||||
{
|
||||
return g_idx * static_cast<long_index_t>(Batchstride_A_);
|
||||
}
|
||||
|
||||
__host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
|
||||
{
|
||||
return g_idx * static_cast<long_index_t>(Batchstride_B_);
|
||||
}
|
||||
|
||||
__host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
|
||||
{
|
||||
const index_t G1 = e_grid_desc_g0_g1_m_n_.GetLength(I1);
|
||||
index_t b0 = g_idx / G1;
|
||||
index_t b1 = g_idx - b0 * G1; // g_idx % G1
|
||||
return e_grid_desc_g0_g1_m_n_.CalculateOffset(make_multi_index(b0, b1, 0, 0));
|
||||
}
|
||||
|
||||
private:
|
||||
index_t Batchstride_A_;
|
||||
index_t Batchstride_B_;
|
||||
EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n_;
|
||||
};
|
||||
|
||||
using GridwiseGemm = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle<
|
||||
ADataType, // TODO: distinguish A/B datatype
|
||||
GemmAccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
AGridDesc_AK0_M_AK1,
|
||||
BGridDesc_BK0_N_BK1,
|
||||
EGridDesc_M_N,
|
||||
NumGemmKPrefetchStage,
|
||||
BlockSize,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
AK1,
|
||||
BK1,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
ABlockTransferThreadClusterArrangeOrder,
|
||||
ABlockTransferSrcAccessOrder,
|
||||
ABlockTransferSrcVectorDim,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
ABlockTransferDstScalarPerVector_AK1,
|
||||
false,
|
||||
ABlockLdsExtraM,
|
||||
BBlockTransferThreadClusterLengths_BK0_N_BK1,
|
||||
BBlockTransferThreadClusterArrangeOrder,
|
||||
BBlockTransferSrcAccessOrder,
|
||||
BBlockTransferSrcVectorDim,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
BBlockTransferDstScalarPerVector_BK1,
|
||||
false,
|
||||
BBlockLdsExtraN,
|
||||
CShuffleMXdlPerWavePerShuffle,
|
||||
CShuffleNXdlPerWavePerShuffle,
|
||||
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
CDEBlockTransferScalarPerVector_NPerBlock,
|
||||
LoopSched>;
|
||||
|
||||
using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = decltype(
|
||||
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}));
|
||||
using Block2CTileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
|
||||
|
||||
// Argument
|
||||
struct Argument : public BaseArgument
|
||||
{
|
||||
Argument(const ADataType* p_a_grid,
|
||||
const BDataType* p_b_grid,
|
||||
EDataType* p_e_grid,
|
||||
index_t M,
|
||||
index_t N,
|
||||
index_t K,
|
||||
index_t stride_A,
|
||||
index_t stride_B,
|
||||
index_t batch_stride_A,
|
||||
index_t batch_stride_B,
|
||||
BatchedGemmCPermuteDesc batched_gemm_c_permute_desc,
|
||||
index_t BatchCount,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation cde_element_op)
|
||||
: p_a_grid_{p_a_grid},
|
||||
p_b_grid_{p_b_grid},
|
||||
p_e_grid_{p_e_grid},
|
||||
BatchCount_(BatchCount),
|
||||
a_grid_desc_ak0_m_ak1_{
|
||||
DeviceBatchedGemmCPermuteXdl::MakeAGridDescriptor_AK0_M_AK1(M, K, stride_A)},
|
||||
b_grid_desc_bk0_n_bk1_{
|
||||
DeviceBatchedGemmCPermuteXdl::MakeBGridDescriptor_BK0_N_BK1(K, N, stride_B)},
|
||||
e_grid_desc_m_n_{DeviceBatchedGemmCPermuteXdl::MakeEGridDescriptor_M_N(
|
||||
batched_gemm_c_permute_desc.M_,
|
||||
batched_gemm_c_permute_desc.N_,
|
||||
batched_gemm_c_permute_desc.stride_M_,
|
||||
batched_gemm_c_permute_desc.stride_N_)},
|
||||
e_grid_desc_g0_g1_m_n_{DeviceBatchedGemmCPermuteXdl::MakeEGridDescriptor_G0_G1_M_N(
|
||||
batched_gemm_c_permute_desc.G0_,
|
||||
batched_gemm_c_permute_desc.G1_,
|
||||
batched_gemm_c_permute_desc.M_,
|
||||
batched_gemm_c_permute_desc.N_,
|
||||
batched_gemm_c_permute_desc.stride_G0_,
|
||||
batched_gemm_c_permute_desc.stride_G1_,
|
||||
batched_gemm_c_permute_desc.stride_M_,
|
||||
batched_gemm_c_permute_desc.stride_N_)},
|
||||
c_grid_desc_mblock_mperblock_nblock_nperblock{},
|
||||
compute_ptr_offset_of_batch_{batch_stride_A, batch_stride_B, e_grid_desc_g0_g1_m_n_},
|
||||
block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
|
||||
a_element_op_{a_element_op},
|
||||
b_element_op_{b_element_op},
|
||||
cde_element_op_{cde_element_op}
|
||||
{
|
||||
|
||||
if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
|
||||
b_grid_desc_bk0_n_bk1_,
|
||||
e_grid_desc_m_n_,
|
||||
block_2_ctile_map_))
|
||||
{
|
||||
c_grid_desc_mblock_mperblock_nblock_nperblock =
|
||||
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
|
||||
e_grid_desc_m_n_);
|
||||
}
|
||||
}
|
||||
|
||||
// private:
|
||||
const ADataType* p_a_grid_;
|
||||
const BDataType* p_b_grid_;
|
||||
EDataType* p_e_grid_;
|
||||
index_t BatchCount_;
|
||||
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
|
||||
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
|
||||
EGridDesc_M_N e_grid_desc_m_n_;
|
||||
EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n_;
|
||||
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock c_grid_desc_mblock_mperblock_nblock_nperblock;
|
||||
ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
|
||||
Block2CTileMap block_2_ctile_map_;
|
||||
AElementwiseOperation a_element_op_;
|
||||
BElementwiseOperation b_element_op_;
|
||||
CDEElementwiseOperation cde_element_op_;
|
||||
};
|
||||
|
||||
// Invoker
|
||||
struct Invoker : public BaseInvoker
|
||||
{
|
||||
using Argument = DeviceBatchedGemmCPermuteXdl::Argument;
|
||||
|
||||
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
|
||||
{
|
||||
{
|
||||
std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
|
||||
<< arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
|
||||
<< arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
|
||||
<< arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
|
||||
|
||||
std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
|
||||
<< arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
|
||||
<< arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
|
||||
<< arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
|
||||
|
||||
std::cout << "arg.e_grid_desc_m_n_{" << arg.e_grid_desc_m_n_.GetLength(I0) << ", "
|
||||
<< arg.e_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
|
||||
}
|
||||
|
||||
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
|
||||
arg.b_grid_desc_bk0_n_bk1_,
|
||||
arg.e_grid_desc_m_n_,
|
||||
arg.block_2_ctile_map_))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! GridwiseBatchedGemmCPermute_km_kn_m0m1n0n1_xdlops_v2r3 has invalid "
|
||||
"setting");
|
||||
}
|
||||
|
||||
const index_t grid_size =
|
||||
arg.block_2_ctile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.BatchCount_;
|
||||
|
||||
const auto K =
|
||||
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
|
||||
|
||||
float ave_time = 0;
|
||||
|
||||
auto launch_kernel = [&](auto has_main_k_block_loop_) {
|
||||
const auto kernel = kernel_batched_gemm_c_permute_xdl<
|
||||
GridwiseGemm,
|
||||
ADataType, // TODO: distiguish A/B datatype
|
||||
EDataType,
|
||||
AGridDesc_AK0_M_AK1,
|
||||
BGridDesc_BK0_N_BK1,
|
||||
typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CDEElementwiseOperation,
|
||||
ComputePtrOffsetOfStridedBatch,
|
||||
remove_reference_t<Block2CTileMap>,
|
||||
has_main_k_block_loop_>;
|
||||
|
||||
return launch_and_time_kernel(stream_config,
|
||||
kernel,
|
||||
dim3(grid_size),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
arg.p_a_grid_,
|
||||
arg.p_b_grid_,
|
||||
arg.p_e_grid_,
|
||||
arg.BatchCount_,
|
||||
arg.a_grid_desc_ak0_m_ak1_,
|
||||
arg.b_grid_desc_bk0_n_bk1_,
|
||||
arg.c_grid_desc_mblock_mperblock_nblock_nperblock,
|
||||
arg.a_element_op_,
|
||||
arg.b_element_op_,
|
||||
arg.cde_element_op_,
|
||||
arg.compute_ptr_offset_of_batch_,
|
||||
arg.block_2_ctile_map_);
|
||||
};
|
||||
|
||||
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
|
||||
{
|
||||
ave_time = launch_kernel(integral_constant<bool, true>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
ave_time = launch_kernel(integral_constant<bool, false>{});
|
||||
}
|
||||
|
||||
return ave_time;
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
float Run(const BaseArgument* p_arg,
|
||||
const StreamConfig& stream_config = StreamConfig{}) override
|
||||
{
|
||||
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
|
||||
}
|
||||
};
|
||||
|
||||
static constexpr bool IsValidCompilationParameter()
|
||||
{
|
||||
// TODO: properly implement this check
|
||||
return true;
|
||||
}
|
||||
|
||||
static bool IsSupportedArgument(const Argument& arg)
|
||||
{
|
||||
return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
|
||||
arg.b_grid_desc_bk0_n_bk1_,
|
||||
arg.e_grid_desc_m_n_,
|
||||
arg.block_2_ctile_map_);
|
||||
}
|
||||
|
||||
static auto MakeArgument(const ADataType* p_a,
|
||||
const BDataType* p_b,
|
||||
EDataType* p_c,
|
||||
index_t M,
|
||||
index_t N,
|
||||
index_t K,
|
||||
index_t stride_A,
|
||||
index_t stride_B,
|
||||
index_t batch_stride_A,
|
||||
index_t batch_stride_B,
|
||||
BatchedGemmCPermuteDesc batched_gemm_c_permute_desc,
|
||||
index_t BatchCount,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation cde_element_op)
|
||||
{
|
||||
return Argument{p_a,
|
||||
p_b,
|
||||
p_c,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
batch_stride_A,
|
||||
batch_stride_B,
|
||||
batched_gemm_c_permute_desc,
|
||||
BatchCount,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op};
|
||||
}
|
||||
|
||||
static auto MakeInvoker() { return Invoker{}; }
|
||||
|
||||
// polymorphic
|
||||
std::unique_ptr<BaseArgument>
|
||||
MakeArgumentPointer(const void* p_a,
|
||||
const void* p_b,
|
||||
void* p_c,
|
||||
index_t M,
|
||||
index_t N,
|
||||
index_t K,
|
||||
index_t stride_A,
|
||||
index_t stride_B,
|
||||
index_t batch_stride_A,
|
||||
index_t batch_stride_B,
|
||||
BatchedGemmCPermuteDesc batched_gemm_c_permute_desc,
|
||||
index_t BatchCount,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CDEElementwiseOperation cde_element_op) override
|
||||
{
|
||||
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
|
||||
static_cast<const BDataType*>(p_b),
|
||||
static_cast<EDataType*>(p_c),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
batch_stride_A,
|
||||
batch_stride_B,
|
||||
batched_gemm_c_permute_desc,
|
||||
BatchCount,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
|
||||
{
|
||||
return std::make_unique<Invoker>(Invoker{});
|
||||
}
|
||||
|
||||
// polymorphic
|
||||
std::string GetTypeString() const override
|
||||
{
|
||||
auto str = std::stringstream();
|
||||
|
||||
// clang-format off
|
||||
str << "DeviceBatchedGemmCPermuteXdl"
|
||||
<< "<"
|
||||
<< BlockSize << ", "
|
||||
<< MPerBlock << ", "
|
||||
<< NPerBlock << ", "
|
||||
<< KPerBlock
|
||||
<< ">";
|
||||
// clang-format on
|
||||
|
||||
return str.str();
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user