Skip to content

GitLab

"projects/vscode:/vscode.git/clone" did not exist on "57fa0803cb143e8d9deb63b9317c21f85bf48684"
Commit 847359c6 authored by Chao Liu's avatar Chao Liu
Browse files

adding output shuffle in conv+bias+relu+add

parent bc6513a2
Expand all Show whitespace changes
Inline Side-by-side
Showing with 3384 additions and 418 deletions
composable_kernel/include/tensor_operation/blockwise_tensor_slice_transfer.hpp → composable_kernel/include/tensor_operation/blockwise_tensor_slice_transfer_v4r1.hpp
View file @ 847359c6
......@@ -5,7 +5,7 @@
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "cluster_descriptor.hpp"
#include "threadwise_tensor_slice_transfer_v3r2.hpp"
#include "threadwise_tensor_slice_transfer_v3r1.hpp"
namespace ck {
......@@ -35,13 +35,13 @@ template <index_t BlockSize,
index_t DstScalarStrideInVector,
bool ThreadTransferSrcResetCoordinateAfterRun,
bool ThreadTransferDstResetCoordinateAfterRun>
struct BlockwiseTensorSliceTransfer_v4
struct BlockwiseTensorSliceTransfer_v4r1
{
static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
using Index = MultiIndex<nDim>;
__device__ constexpr BlockwiseTensorSliceTransfer_v4(
__device__ constexpr BlockwiseTensorSliceTransfer_v4r1(
const SrcDesc& src_desc,
const Index& src_block_slice_origin,
const SrcElementwiseOperation& src_element_op,
......@@ -165,7 +165,7 @@ struct BlockwiseTensorSliceTransfer_v4
make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
using ThreadwiseTransfer =
ThreadwiseTensorSliceTransfer_v3r2<ThreadSliceLengths,
ThreadwiseTensorSliceTransfer_v3r1<ThreadSliceLengths,
SrcElementwiseOperation,
DstElementwiseOperation,
DstInMemOp,
......
composable_kernel/include/tensor_operation/blockwise_tensor_slice_transfer_v4r3.hpp 0 → 100644
View file @ 847359c6
#ifndef CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V4R3_HPP
#define CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V4R3_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "cluster_descriptor.hpp"
#include "threadwise_tensor_slice_transfer_v3r3.hpp"
namespace ck {
// this version does following things to avoid scratch memory issue
// 1. Use StaticallyIndexedArray instead of C array for thread buffer
// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
template <index_t BlockSize,
typename SrcElementwiseOperation,
typename DstElementwiseOperation,
InMemoryDataOperationEnum_t DstInMemOp,
typename BlockSliceLengths,
typename ThreadSliceLengths,
typename ThreadClusterLengths,
typename ThreadClusterArrangeOrder,
typename SrcData,
typename DstData,
typename SrcDesc,
typename DstDesc,
typename Dst0Desc, // this is really one of sources, but it has same shape as DstDesc
typename Dst1Desc, // this is really one of sources, but it has same shape as DstDesc
typename SrcDimAccessOrder,
typename DstDimAccessOrder,
index_t SrcVectorDim,
index_t DstVectorDim,
index_t SrcScalarPerVector,
index_t DstScalarPerVector,
index_t SrcScalarStrideInVector,
index_t DstScalarStrideInVector,
bool ThreadTransferSrcResetCoordinateAfterRun,
bool ThreadTransferDstResetCoordinateAfterRun>
struct BlockwiseTensorSliceTransfer_v4r3
{
static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
using Index = MultiIndex<nDim>;
__device__ constexpr BlockwiseTensorSliceTransfer_v4r3(
const SrcDesc& src_desc,
const Index& src_block_slice_origin,
const SrcElementwiseOperation& src_element_op,
const DstDesc& dst_desc,
const Dst0Desc& dst0_desc,
const Dst1Desc& dst1_desc,
const Index& dst_block_slice_origin,
const DstElementwiseOperation& dst_element_op)
: threadwise_transfer_(src_desc,
make_zero_multi_index<nDim>(),
src_element_op,
dst_desc,
dst0_desc,
dst1_desc,
make_zero_multi_index<nDim>(),
dst_element_op)
{
static_assert(nDim == remove_reference_t<remove_cv_t<SrcDesc>>::GetNumOfDimension() &&
nDim == remove_reference_t<remove_cv_t<DstDesc>>::GetNumOfDimension() &&
nDim == remove_reference_t<remove_cv_t<Dst0Desc>>::GetNumOfDimension() &&
nDim == remove_reference_t<remove_cv_t<Dst1Desc>>::GetNumOfDimension() &&
nDim == BlockSliceLengths::Size() && nDim == ThreadSliceLengths::Size() &&
nDim == ThreadClusterLengths::Size() &&
nDim == ThreadClusterArrangeOrder::Size() &&
nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
"wrong! nDim not consistent");
static_assert(
is_same<BlockSliceLengths, decltype(ThreadSliceLengths{} * ThreadClusterLengths{})>{},
"wrong! threads should be mapped to cover entire slicing window");
static_assert(BlockSize >= thread_cluster_desc_.GetElementSize(),
"wrong! BlockSize too small");
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
{
const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
make_multi_index(get_thread_local_1d_id()));
const auto thread_data_idx_begin = thread_cluster_idx * ThreadSliceLengths{};
threadwise_transfer_.SetSrcSliceOrigin(src_desc,
src_block_slice_origin + thread_data_idx_begin);
threadwise_transfer_.SetDstSliceOrigin(
dst_desc, dst0_desc, dst1_desc, dst_block_slice_origin + thread_data_idx_begin);
}
}
template <typename SrcBuffer>
__device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.RunRead(src_desc, src_buf);
}
}
// this is really load dst0 and dst1 and write to dst
template <typename DstBuffer, typename Dst0Bufferm typename Dst1Buffer>
__device__ void RunWrite(const DstDesc& dst_desc,
DstBuffer& dst_buf,
const Dst0Desc& dst0_desc,
const Dst0Buffer& dst0_buf,
const Dst1Desc& dst1_desc,
const Dst1Buffer& dst1_buf)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.RunWrite(
dst_desc, dst_buf, dst0_desc, dst0_buf, dst1_desc, dst1_buf);
}
}
template <typename SrcBuffer, typename DstBuffer>
__device__ void Run(const SrcDesc& src_desc,
const SrcBuffer& src_buf,
const DstDesc& dst_desc,
DstBuffer& dst_buf,
const Dst0Desc& dst0_desc,
const Dst0Buffer& dst0_buf,
const Dst1Desc& dst1_desc,
const Dst1Buffer& dst1_buf);
{
RunRead(src_desc, src_buf);
RunWrite(dst_desc, dst_buf, dst0_desc, dst0_buf, dst1_desc, dst1_buf);
}
__device__ void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& step)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrcSliceWindow(src_desc, step);
}
}
__device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
const Dst0Desc& dst0_desc,
const Dst1Desc& dst1_desc,
const Index& step)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveDstSliceWindow(dst_desc, dst0_desc, dst1_desc, step);
}
}
private:
static constexpr auto thread_cluster_desc_ =
make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
using ThreadwiseTransfer =
ThreadwiseTensorSliceTransfer_v3r3<ThreadSliceLengths,
SrcElementwiseOperation,
DstElementwiseOperation,
DstInMemOp,
SrcData,
DstData,
SrcDesc,
DstDesc,
Dst0Desc,
Dst1Desc,
SrcDimAccessOrder,
DstDimAccessOrder,
SrcVectorDim,
DstVectorDim,
SrcScalarPerVector,
DstScalarPerVector,
SrcScalarStrideInVector,
DstScalarStrideInVector,
ThreadTransferSrcResetCoordinateAfterRun,
ThreadTransferDstResetCoordinateAfterRun>;
ThreadwiseTransfer threadwise_transfer_;
};
} // namespace ck
#endif
composable_kernel/include/tensor_operation/blockwise_tensor_slice_transfer_v2.hpp → composable_kernel/include/tensor_operation/blockwise_tensor_slice_transfer_v5r1.hpp
View file @ 847359c6
#ifndef CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V2_HPP
#define CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V2_HPP
#ifndef CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V5R1_HPP
#define CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V5R1_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "cluster_descriptor.hpp"
#include "threadwise_tensor_slice_transfer_v2.hpp"
#include "threadwise_tensor_slice_transfer_v5r1.hpp"
namespace ck {
......@@ -31,13 +31,13 @@ template <index_t BlockSize,
typename DstVectorTensorContiguousDimOrder,
bool ThreadTransferSrcResetCoordinateAfterRun,
bool ThreadTransferDstResetCoordinateAfterRun>
struct BlockwiseTensorSliceTransfer_v4r1
struct BlockwiseTensorSliceTransfer_v5r1
{
static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
using Index = MultiIndex<nDim>;
__device__ constexpr BlockwiseTensorSliceTransfer_v4r1(const SrcDesc& src_desc,
__device__ constexpr BlockwiseTensorSliceTransfer_v5r1(const SrcDesc& src_desc,
const Index& src_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin)
......@@ -134,7 +134,7 @@ struct BlockwiseTensorSliceTransfer_v4r1
make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
using ThreadwiseTransfer =
ThreadwiseTensorSliceTransfer_v3r1<ThreadSliceLengths,
ThreadwiseTensorSliceTransfer_v5r1<ThreadSliceLengths,
DstInMemOp,
SrcData,
DstData,
......
composable_kernel/include/tensor_operation/gridwise_contraction_dlops_v1r2.hpp
View file @ 847359c6
......@@ -381,7 +381,7 @@ struct GridwiseContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0_GM1_GN
"wrong!");
// A matrix blockwise copy
auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v4r1<
auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
BlockSize,
InMemoryDataOperationEnum_t::Set,
Sequence<GK0PerBlock, GM0, 1, GM1PerBlockGM11, GK1.value>,
......@@ -405,7 +405,7 @@ struct GridwiseContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0_GM1_GN
make_multi_index(0, 0, 0, 0, 0));
// B matrix blockwise copy
auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v4r1<
auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
BlockSize,
InMemoryDataOperationEnum_t::Set,
Sequence<GK0PerBlock, GN0, 1, GN1PerBlockGN11, GK1.value>,
......
composable_kernel/include/tensor_operation/gridwise_gemm_dlops_v1r3.hpp
View file @ 847359c6
......@@ -6,7 +6,7 @@
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "blockwise_gemm_dlops_v2r3.hpp"
#include "blockwise_tensor_slice_transfer_v2.hpp"
#include "blockwise_tensor_slice_transfer_v5r1.hpp"
#include "threadwise_tensor_slice_transfer_v2.hpp"
#include "threadwise_tensor_slice_set.hpp"
......@@ -380,7 +380,7 @@ struct GridwiseGemmDlops_km_kn_mn_v1r3
"wrong!");
// A matrix blockwise copy
auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v4r1<
auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
BlockSize,
InMemoryDataOperationEnum_t::Set,
Sequence<KPerBlock, 1, MPerBlockM1, K1.value>,
......@@ -404,7 +404,7 @@ struct GridwiseGemmDlops_km_kn_mn_v1r3
make_multi_index(0, 0, 0, 0));
// B matrix blockwise copy
auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v4r1<
auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
BlockSize,
InMemoryDataOperationEnum_t::Set,
Sequence<KPerBlock, 1, NPerBlockN1, K1.value>,
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r3.hpp
View file @ 847359c6
......@@ -6,9 +6,8 @@
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "blockwise_gemm_xdlops.hpp"
#include "blockwise_tensor_slice_transfer.hpp"
#include "blockwise_tensor_slice_transfer_v4r1.hpp"
#include "threadwise_tensor_slice_transfer.hpp"
#include "threadwise_tensor_slice_set.hpp"
namespace ck {
......@@ -435,7 +434,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
// A matrix blockwise copy
auto a_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
......@@ -466,7 +465,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
// B matrix blockwise copy
auto b_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r4.hpp
View file @ 847359c6
......@@ -6,9 +6,8 @@
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "blockwise_gemm_xdlops.hpp"
#include "blockwise_tensor_slice_transfer.hpp"
#include "blockwise_tensor_slice_transfer_v4r1.hpp"
#include "threadwise_tensor_slice_transfer.hpp"
#include "threadwise_tensor_slice_set.hpp"
namespace ck {
......@@ -452,7 +451,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
}();
// A matrix blockwise copy
auto a_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
......@@ -483,7 +482,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
// B matrix blockwise copy
auto b_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r5.hpp
View file @ 847359c6
......@@ -6,9 +6,8 @@
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "blockwise_gemm_xdlops.hpp"
#include "blockwise_tensor_slice_transfer.hpp"
#include "blockwise_tensor_slice_transfer_v4r1.hpp"
#include "threadwise_tensor_slice_transfer_v1r4.hpp"
#include "threadwise_tensor_slice_set.hpp"
namespace ck {
......@@ -405,7 +404,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r5
// A matrix blockwise copy
auto a_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
......@@ -436,7 +435,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r5
// B matrix blockwise copy
auto b_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r6.hpp
View file @ 847359c6
......@@ -6,9 +6,8 @@
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "blockwise_gemm_xdlops.hpp"
#include "blockwise_tensor_slice_transfer.hpp"
#include "blockwise_tensor_slice_transfer_v4r1.hpp"
#include "threadwise_tensor_slice_transfer_v1r5.hpp"
#include "threadwise_tensor_slice_set.hpp"
namespace ck {
......@@ -391,7 +390,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r6
// A matrix blockwise copy
auto a_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
......@@ -422,7 +421,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r6
// B matrix blockwise copy
auto b_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v3r1.hpp
View file @ 847359c6
......@@ -6,9 +6,8 @@
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "blockwise_gemm_xdlops.hpp"
#include "blockwise_tensor_slice_transfer.hpp"
#include "blockwise_tensor_slice_transfer_v4r1.hpp"
#include "threadwise_tensor_slice_transfer.hpp"
#include "threadwise_tensor_slice_set.hpp"
namespace ck {
......@@ -111,6 +110,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
// K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{};
// TODO: need to calculate LDS usage for C shuffle
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
constexpr auto max_lds_align = K1;
......@@ -354,7 +354,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
// A matrix blockwise copy
auto a_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
......@@ -385,7 +385,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
// B matrix blockwise copy
auto b_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
......@@ -654,7 +654,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
auto c_block_copy_lds_to_global = BlockwiseTensorSliceTransfer_v4<
auto c_block_copy_lds_to_global = BlockwiseTensorSliceTransfer_v4r1<
BlockSize, // index_t BlockSize,
ck::tensor_operation::element_wise::PassThrough, // SrcElementwiseOperation,
CElementwiseOperation, // DstElementwiseOperation,
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v1r4.hpp
View file @ 847359c6
......@@ -141,7 +141,8 @@ struct ThreadwiseTensorSliceTransfer_v1r4
Number<nDim>{});
// make forward steps: dst0
// WARNING!!!!!!: this logic is only correct if DstScalarPerVector=1
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst0_forward_steps = generate_tuple(
[&](auto i) {
......@@ -157,7 +158,8 @@ struct ThreadwiseTensorSliceTransfer_v1r4
Number<nDim>{});
// make forward steps: dst1
// WARNING!!!!!!: this logic is only correct if DstScalarPerVector=1
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst1_forward_steps = generate_tuple(
[&](auto i) {
......@@ -187,7 +189,8 @@ struct ThreadwiseTensorSliceTransfer_v1r4
Number<nDim>{});
// make backward steps: dst0
// WARNING!!!!!!: this logic is only correct if DstScalarPerVector=1
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst0_backward_steps = generate_tuple(
[&](auto i) {
......@@ -203,7 +206,8 @@ struct ThreadwiseTensorSliceTransfer_v1r4
Number<nDim>{});
// make backward steps: dst1
// WARNING!!!!!!: this logic is only correct if DstScalarPerVector=1
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst1_backward_steps = generate_tuple(
[&](auto i) {
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v3r2.hpp → composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v3r3.hpp
View file @ 847359c6
#ifndef CK_THREADWISE_TENSOR_SLICE_TRANSFER_V3R2_HPP
#define CK_THREADWISE_TENSOR_SLICE_TRANSFER_V3R2_HPP
#ifndef CK_THREADWISE_TENSOR_SLICE_TRANSFER_V3R3_HPP
#define CK_THREADWISE_TENSOR_SLICE_TRANSFER_V3R3_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
......@@ -53,6 +53,8 @@ template <typename SliceLengths,
typename DstData,
typename SrcDesc,
typename DstDesc,
typename Dst0Desc,
typename Dst1Desc,
typename SrcDimAccessOrder,
typename DstDimAccessOrder,
index_t SrcVectorDim,
......@@ -67,26 +69,34 @@ template <typename SliceLengths,
bool DstResetCoordinateAfterRun> // control whether to move back dst coordinate after each
// RunWrite(), will be fused with MoveDstSliceWindow to
// save addr computation
struct ThreadwiseTensorSliceTransfer_v3r2
struct ThreadwiseTensorSliceTransfer_v3r3
{
static constexpr index_t nDim = SliceLengths::Size();
using Index = MultiIndex<nDim>;
using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
using Dst0Coord = decltype(make_tensor_coordinate(Dst0Desc{}, Index{}));
using Dst1Coord = decltype(make_tensor_coordinate(Dst1Desc{}, Index{}));
using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
using Dst0CoordStep = decltype(make_tensor_coordinate_step(Dst0Desc{}, Index{}));
using Dst1CoordStep = decltype(make_tensor_coordinate_step(Dst1Desc{}, Index{}));
__device__ constexpr ThreadwiseTensorSliceTransfer_v3r2(
__device__ constexpr ThreadwiseTensorSliceTransfer_v3r3(
const SrcDesc& src_desc,
const Index& src_slice_origin,
const SrcElementwiseOperation& src_element_op,
const DstDesc& dst_desc,
const Dst0Desc& dst0_desc,
const Dst1Desc& dst1_desc,
const Index& dst_slice_origin,
const DstElementwiseOperation& dst_element_op)
: src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
dst0_coord_(make_tensor_coordinate(dst0_desc, dst_slice_origin)),
dst1_coord_(make_tensor_coordinate(dst1_desc, dst_slice_origin)),
src_element_op_(src_element_op),
dst_element_op_(dst_element_op)
{
......@@ -97,14 +107,18 @@ struct ThreadwiseTensorSliceTransfer_v3r2
src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
}
__device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
__device__ void SetDstSliceOrigin(const DstDesc& dst_desc,
const Dst0Desc& dst0_desc,
const Dst1Desc& dst1_desc,
const Index& dst_slice_origin_idx)
{
dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
dst0_coord_ = make_tensor_coordinate(dst0_desc, dst_slice_origin_idx);
dst1_coord_ = make_tensor_coordinate(dst1_desc, dst_slice_origin_idx);
}
template <typename SrcBuffer, typename SrcStepHacks>
__device__ void
RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf, const SrcStepHacks& src_step_hacks)
template <typename SrcBuffer>
__device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
{
static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Global or
SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
......@@ -138,8 +152,7 @@ struct ThreadwiseTensorSliceTransfer_v3r2
forward_step_idx(j) = (i.value == j.value) ? src_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
src_desc, forward_step_idx, src_step_hacks[I0][i]);
return make_tensor_coordinate_step(src_desc, forward_step_idx);
},
Number<nDim>{});
......@@ -152,8 +165,7 @@ struct ThreadwiseTensorSliceTransfer_v3r2
backward_step_idx(j) = (i.value == j.value) ? -src_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
src_desc, backward_step_idx, src_step_hacks[I1][i]);
return make_tensor_coordinate_step(src_desc, backward_step_idx);
},
Number<nDim>{});
......@@ -346,9 +358,13 @@ struct ThreadwiseTensorSliceTransfer_v3r2
#endif
}
template <typename DstBuffer, typename DstStepHacks>
__device__ void
RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf, const DstStepHacks& dst_step_hacks)
template <typename DstBuffer, typename Dst0Buffer, typename Dst1Buffer>
__device__ void RunWrite(const DstDesc& dst_desc,
DstBuffer& dst_buf,
const Dst0Desc& dst0_desc,
const Dst0Buffer& dst0_buf,
const Dst1Desc& dst1_desc,
const Dst1Buffer& dst1_buf)
{
// if there is transpose, it's done here
// TODO move this elsewhere
......@@ -386,8 +402,39 @@ struct ThreadwiseTensorSliceTransfer_v3r2
forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst_desc, forward_step_idx, dst_step_hacks[I0][i]);
return make_tensor_coordinate_step(dst_desc, forward_step_idx);
},
Number<nDim>{});
// make forward steps: dst0
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst0_forward_steps = generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(dst0_desc, forward_step_idx);
},
Number<nDim>{});
// make forward steps: dst1
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst1_forward_steps = generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(dst1_desc, forward_step_idx);
},
Number<nDim>{});
......@@ -400,8 +447,39 @@ struct ThreadwiseTensorSliceTransfer_v3r2
backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst_desc, backward_step_idx, dst_step_hacks[I1][i]);
return make_tensor_coordinate_step(dst_desc, backward_step_idx);
},
Number<nDim>{});
// make backward steps: dst0
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst0_backward_steps = generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(dst0_desc, backward_step_idx);
},
Number<nDim>{});
// make backward steps: dst1
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst1_backward_steps = generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(dst1_desc, backward_step_idx);
},
Number<nDim>{});
......@@ -511,35 +589,6 @@ struct ThreadwiseTensorSliceTransfer_v3r2
}
}
template <typename SrcBuffer>
__device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
{
constexpr index_t ntransform_src = remove_cvref_t<SrcDesc>::GetNumOfTransform();
constexpr auto zeros = typename uniform_sequence_gen<ntransform_src, 0>::type{};
constexpr auto src_step_hacks =
make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
RunRead(src_desc, src_buf, src_step_hacks);
}
template <typename DstBuffer>
__device__ void RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf)
{
// TODO: why need remove_cvref_t ?
constexpr index_t ntransform_dst = remove_cvref_t<DstDesc>::GetNumOfTransform();
constexpr auto zeros = typename uniform_sequence_gen<ntransform_dst, 0>::type{};
constexpr auto dst_step_hacks =
make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
RunWrite(dst_desc, dst_buf, dst_step_hacks);
}
__device__ static constexpr auto GetSrcCoordinateResetStep()
{
constexpr auto I0 = Number<0>{};
......@@ -678,11 +727,8 @@ struct ThreadwiseTensorSliceTransfer_v3r2
}
// src_slice_origin_step_idx need to be known at compile-time, for performance reason
template <typename SrcMoveSliceWindowStepHack>
__device__ void
MoveSrcSliceWindow(const SrcDesc& src_desc,
const Index& src_slice_origin_step_idx,
const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
__device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
const Index& src_slice_origin_step_idx)
{
// if src coord was not reset by RunRead(), then need to adjust the step here
const auto adjusted_step_idx =
......@@ -690,14 +736,15 @@ struct ThreadwiseTensorSliceTransfer_v3r2
: src_slice_origin_step_idx + GetSrcCoordinateResetStep();
// is it OK to construct a new step every time?
const auto adjusted_step = make_tensor_coordinate_step(
src_desc, adjusted_step_idx, src_move_slice_window_step_hack);
const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
}
// dst_slice_origin_step_idx need to be known at compile-time, for performance reason
__device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
const Dst0Desc dst0_desc,
const Dst1Desc dst1_desc,
const Index& dst_slice_origin_step_idx)
{
// if dst coord was not reset by RunWrite(), then need to adjust the step here
......@@ -709,6 +756,8 @@ struct ThreadwiseTensorSliceTransfer_v3r2
const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
move_tensor_coordinate(dst0_desc, dst0_coord_, adjusted_step);
move_tensor_coordinate(dst1_desc, dst1_coord_, adjusted_step);
}
__device__ static constexpr auto GetSrcThreadScratchDescriptor()
......
example/6_conv2d_fwd_xdl_output_shuffle_bias_relu_add/README.md 0 → 100644
View file @ 847359c6
# Instructions for ```conv_xdl_bias_relu_add``` Example
## Docker script
```bash
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
```
## Build ```conv_xdl_bias_relu_add```
```bash
mkdir build && cd build
```
```bash
# Need to specify target ID, example below is gfx908
cmake \
-D BUILD_DEV=OFF \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
..
```
```bash
make -j conv_xdl_bias_relu_add
```
## Run ```conv_xdl_bias_relu_add```
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1)
#arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
./example/conv_xdl_bias_relu_add 0 1 5
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
wei_k_c_y_x: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
bias_k: dim 1, lengths {256}, strides {1}
resi_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
arg.a_grid_desc_k0_m_k1_{216, 165888, 8}
arg.b_grid_desc_k0_n_k1_{216, 256, 8}
arg.c_grid_desc_m_n_{ 165888, 256}
arg.c0_grid_desc_m_n_{ 165888, 256}
arg.c1_grid_desc_m_n_{ 165888, 256}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 1.71779 ms, 85.4396 TFlops, 194.2 GB/s
```
example/6_conv2d_fwd_xdl_output_shuffle_bias_relu_add/conv2d_fwd_xdl_output_shuffle_bias_relu_add_output.cpp 0 → 100644
View file @ 847359c6
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "tensor_layout.hpp"
#include "device_conv2d_fwd_xdl_output_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp"
#include "element_wise_operation.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InLayout = ck::tensor_layout::convolution::NHWC;
using WeiLayout = ck::tensor_layout::convolution::KYXC;
using OutLayout = ck::tensor_layout::convolution::NHWK;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
// clang-format off
using DeviceConvFwdInstance = ck::tensor_operation::device::
DeviceConv2dFwdXdl_Ouput_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
// | InData| WeiData| OutData| AccData| In| Wei| Out| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
// | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| SrcDstVectorDim| DstScalar| AddExtraM| AddExtraN|
// | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerVector| | |
// | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
<InDataType, WeiDataType, OutDataType, AccDataType, InElementOp, WeiElementOp, OutElementOp, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 4, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>;
// clang-format on
template <typename TIn,
typename TWei,
typename TOut,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp>
void host_reference_calculation(const Tensor<TIn>& in_n_c_hi_wi,
const Tensor<TWei>& wei_k_c_y_x,
Tensor<TOut>& out_n_k_ho_wo,
const Tensor<TOut>& bias_k,
const Tensor<TOut>& resi_n_k_ho_wo,
const std::vector<ck::index_t>& conv_strides,
const std::vector<ck::index_t>& conv_dilations,
const std::vector<ck::index_t>& in_left_pads,
const std::vector<ck::index_t>& /* in_right_pads */,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
{
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
double v = 0;
for(int c = 0; c < wei_k_c_y_x.mDesc.GetLengths()[1]; ++c)
{
for(int y = 0; y < wei_k_c_y_x.mDesc.GetLengths()[2]; ++y)
{
int hi = ho * conv_strides[0] + y * conv_dilations[0] - in_left_pads[0];
for(int x = 0; x < wei_k_c_y_x.mDesc.GetLengths()[3]; ++x)
{
int wi = wo * conv_strides[1] + x * conv_dilations[1] - in_left_pads[1];
if(hi >= 0 && hi < in_n_c_hi_wi.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in_n_c_hi_wi.mDesc.GetLengths()[3])
{
v += in_element_op(static_cast<const double>(in_n_c_hi_wi(n, c, hi, wi))) *
wei_element_op(static_cast<const double>(wei_k_c_y_x(k, c, y, x)));
}
}
}
}
out_n_k_ho_wo(n, k, ho, wo) = out_element_op(v, bias_k(k), resi_n_k_ho_wo(n, k, ho, wo));
};
make_ParallelTensorFunctor(f_nchw,
out_n_k_ho_wo.mDesc.GetLengths()[0],
out_n_k_ho_wo.mDesc.GetLengths()[1],
out_n_k_ho_wo.mDesc.GetLengths()[2],
out_n_k_ho_wo.mDesc.GetLengths()[3])(
std::thread::hardware_concurrency());
}
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
// Conv shape
ck::index_t N = 128;
ck::index_t K = 256;
ck::index_t C = 192;
ck::index_t Y = 3;
ck::index_t X = 3;
ck::index_t Hi = 71;
ck::index_t Wi = 71;
ck::index_t conv_stride_h = 2;
ck::index_t conv_stride_w = 2;
ck::index_t conv_dilation_h = 1;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
}
else if(argc == 19)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
N = std::stoi(argv[4]);
K = std::stoi(argv[5]);
C = std::stoi(argv[6]);
Y = std::stoi(argv[7]);
X = std::stoi(argv[8]);
Hi = std::stoi(argv[9]);
Wi = std::stoi(argv[10]);
conv_stride_h = std::stoi(argv[11]);
conv_stride_w = std::stoi(argv[12]);
conv_dilation_h = std::stoi(argv[13]);
conv_dilation_w = std::stoi(argv[14]);
in_left_pad_h = std::stoi(argv[15]);
in_left_pad_w = std::stoi(argv[16]);
in_right_pad_h = std::stoi(argv[17]);
in_right_pad_w = std::stoi(argv[18]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
}
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
// tensor layout
auto f_host_tensor_descriptor = [](std::size_t N_,
std::size_t C_,
std::size_t H,
std::size_t W,
auto layout) {
if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
ck::is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
ck::is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWC>::value ||
ck::is_same<decltype(layout),
ck::tensor_layout::convolution::KYXC>::value ||
ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<OutDataType> out_n_k_ho_wo_host_result(
f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
Tensor<OutDataType> out_n_k_ho_wo_device_result(
f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
// bias: assume contiguous 1d vector
Tensor<OutDataType> bias_k(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(K)})));
// residual: assume same layout as output tensor
Tensor<OutDataType> resi_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo_host_result.mDesc << std::endl;
std::cout << "bias_k: " << bias_k.mDesc << std::endl;
std::cout << "resi_n_k_ho_wo: " << resi_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
bias_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
resi_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break;
default:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
bias_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
resi_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) *
out_n_k_ho_wo_device_result.mDesc.GetElementSpace());
DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpace());
DeviceMem resi_device_buf(sizeof(OutDataType) * resi_n_k_ho_wo.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
bias_device_buf.ToDevice(bias_k.mData.data());
resi_device_buf.ToDevice(resi_n_k_ho_wo.mData.data());
auto conv = DeviceConvFwdInstance{};
auto invoker = conv.MakeInvoker();
auto argument =
conv.MakeArgument(static_cast<const InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<const WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
static_cast<const OutDataType*>(resi_device_buf.GetDeviceBuffer()),
N,
K,
C,
std::vector<ck::index_t>{{Hi, Wi}},
std::vector<ck::index_t>{{Y, X}},
std::vector<ck::index_t>{{Ho, Wo}},
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
if(!conv.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device operator with the specified compilation parameters does "
"not support this problem");
}
float ave_time = invoker.Run(argument, nrepeat);
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo) + sizeof(OutDataType) * (K) +
sizeof(OutDataType) * (N * K * Ho * Wo);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
if(do_verification)
{
host_reference_calculation(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
bias_k,
resi_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
check_error(out_n_k_ho_wo_host_result, out_n_k_ho_wo_device_result);
}
}
example/6_conv2d_fwd_xdl_output_shuffle_bias_relu_add/include/device_conv_fwd_xdl_bias_activation_add.hpp 0 → 100644
View file @ 847359c6
#ifndef DEVICE_CONV_FWD_XDL_BIAS_ACTIVATION_ADD_HPP
#define DEVICE_CONV_FWD_XDL_BIAS_ACTIVATION_ADD_HPP
#include <iostream>
#include "device.hpp"
#include "device_base.hpp"
#include "device_conv.hpp"
#include "common_header.hpp"
#include "tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_xdlops_v2r3.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <ck::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout,
typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t K0PerBlock,
ck::index_t K1,
ck::index_t MPerXDL,
ck::index_t NPerXDL,
ck::index_t MXdlPerWave,
ck::index_t NXdlPerWave,
typename ABlockTransferThreadSliceLengths_K0_M_K1,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
ck::index_t ABlockTransferSrcVectorDim,
ck::index_t ABlockTransferSrcScalarPerVector,
ck::index_t ABlockTransferDstScalarPerVector_K1,
typename BBlockTransferThreadSliceLengths_K0_N_K1,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
ck::index_t BBlockTransferSrcVectorDim,
ck::index_t BBlockTransferSrcScalarPerVector,
ck::index_t BBlockTransferDstScalarPerVector_K1,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector,
bool ABlockLdsAddExtraM,
bool BBlockLdsAddExtraN>
struct DeviceConvFwdXdl_bias_activation_add;
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
example/CMakeLists.txt
View file @ 847359c6
......@@ -17,6 +17,7 @@ set(CONV2D_FWD_XDL_SOURCE 4_conv2d_fwd_xdl/conv2d_fwd_xdl.cpp)
set(CONV2D_FWD_XDL_OUTPUT_SHUFFLE_SOURCE 4_conv2d_fwd_xdl_output_shuffle/conv2d_fwd_xdl_output_shuffle.cpp)
set(CONV2D_FWD_XDL_BIAS_RELU_SOURCE 5_conv2d_fwd_xdl_bias_relu/conv2d_fwd_xdl_bias_relu.cpp)
set(CONV2D_FWD_XDL_BIAS_RELU_ADD_SOURCE 6_conv2d_fwd_xdl_bias_relu_add/conv2d_fwd_xdl_bias_relu_add.cpp)
set(CONV2D_FWD_XDL_OUTPUT_SHUFFLE_BIAS_RELU_ADD_SOURCE 6_conv2d_fwd_xdl_output_shuffle_bias_relu_add/conv2d_fwd_xdl_bias_relu_add.cpp)
add_executable(gemm_xdl ${GEMM_XDL_SOURCE})
add_executable(gemm_xdl_bias_relu_add ${GEMM_XDL_BIAS_RELU_ADD_SOURCE})
......@@ -24,6 +25,7 @@ add_executable(conv2d_fwd_xdl ${CONV2D_FWD_XDL_SOURCE})
add_executable(conv2d_fwd_xdl_output_shuffle ${CONV2D_FWD_XDL_OUTPUT_SHUFFLE_SOURCE})
add_executable(conv2d_fwd_xdl_bias_relu ${CONV2D_FWD_XDL_BIAS_RELU_SOURCE})
add_executable(conv2d_fwd_xdl_bias_relu_add ${CONV2D_FWD_XDL_BIAS_RELU_ADD_SOURCE})
add_executable(conv2d_fwd_xdl_output_shuffle_bias_relu_add ${CONV2D_FWD_XDL_OUTPUT_SHUFFLE_BIAS_RELU_ADD_SOURCE})
target_link_libraries(gemm_xdl PRIVATE host_tensor)
target_link_libraries(gemm_xdl_bias_relu_add PRIVATE host_tensor)
......@@ -31,3 +33,4 @@ target_link_libraries(conv2d_fwd_xdl PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl_output_shuffle PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl_bias_relu PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl_bias_relu_add PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl_output_shuffle_bias_relu_add PRIVATE host_tensor)
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