Skip to content

GitLab

Commit f9c478e2 authored by ltqin's avatar ltqin
Browse files

Merge branch 'develop' into bmatrix_skip_lds

parents 7d85d04a 91d8b7d6
Hide whitespace changes
Inline Side-by-side
Showing with 1103 additions and 604 deletions
include/ck/stream_config.hpp 0 → 100644
View file @ f9c478e2
#pragma once
#include <hip/hip_runtime.h>
#include <hip/hip_fp16.h>
struct StreamConfig
{
hipStream_t stream_id_ = nullptr;
bool time_kernel_ = false;
};
include/ck/tensor_description/tensor_descriptor_helper.hpp
View file @ f9c478e2
#ifndef CK_TENSOR_DESCRIPTOR_HELPER_HPP
#define CK_TENSOR_DESCRIPTOR_HELPER_HPP
#pragma once
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "multi_index_transform_helper.hpp"
......@@ -35,6 +33,12 @@ __host__ __device__ constexpr auto calculate_element_space_size_impl(const Lengt
}
#endif
// Lengths..., Strides... could be:
// 1) index_t, which is known at run-time, or
// 2) Number<>, which is known at compile-time
// element_space_size could be:
// 1) long_index_t, or
// 2) LongNumber<>
template <typename... Lengths,
typename... Strides,
typename enable_if<sizeof...(Lengths) == sizeof...(Strides), bool>::type = false>
......@@ -68,10 +72,10 @@ __host__ __device__ constexpr auto make_naive_tensor_descriptor(const Tuple<Leng
}
};
const auto element_space_size = f(f, Number<0>{}, Number<1>{});
const auto element_space_size = f(f, Number<0>{}, LongNumber<1>{});
#else
const auto element_space_size =
calculate_element_space_size_impl(lengths, strides, Number<0>{}, Number<1>{});
calculate_element_space_size_impl(lengths, strides, Number<0>{}, LongNumber<1>{});
#endif
return TensorDescriptor<remove_cv_t<decltype(transforms)>,
......@@ -82,9 +86,12 @@ __host__ __device__ constexpr auto make_naive_tensor_descriptor(const Tuple<Leng
element_space_size};
}
// Lengths... can be:
// 1) index_t, which is known at run-time
// Lengths... could be:
// 1) index_t, which is known at run-time, or
// 2) Number<>, which is known at compile-time
// element_space_size could be:
// 1) long_index_t, or
// 2) LongNumber<>
template <typename... Lengths>
__host__ __device__ constexpr auto
make_naive_tensor_descriptor_packed(const Tuple<Lengths...>& lengths)
......@@ -100,7 +107,7 @@ make_naive_tensor_descriptor_packed(const Tuple<Lengths...>& lengths)
constexpr auto visible_dim_hidden_ids = typename arithmetic_sequence_gen<1, N + 1, 1>::type{};
const auto element_space_size = container_reduce(lengths, math::multiplies{}, Number<1>{});
const auto element_space_size = container_reduce(lengths, math::multiplies{}, LongNumber<1>{});
return TensorDescriptor<remove_cv_t<decltype(transforms)>,
remove_cv_t<decltype(low_dim_hidden_idss)>,
......@@ -110,6 +117,12 @@ make_naive_tensor_descriptor_packed(const Tuple<Lengths...>& lengths)
element_space_size};
}
// Lengths... could be:
// 1) index_t, which is known at run-time, or
// 2) Number<>, which is known at compile-time
// align could be:
// 1) index_t, or
// 2) Number<>
template <typename... Lengths, typename Align>
__host__ __device__ constexpr auto
make_naive_tensor_descriptor_aligned(const Tuple<Lengths...>& lengths, Align align)
......@@ -146,4 +159,3 @@ make_naive_tensor_descriptor_aligned(const Tuple<Lengths...>& lengths, Align ali
}
} // namespace ck
#endif
include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r3.hpp → include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp
View file @ f9c478e2
#ifndef CK_BLOCKWISE_GEMM_DLOPS_V2R3_HPP
#define CK_BLOCKWISE_GEMM_DLOPS_V2R3_HPP
#pragma once
#include "common_header.hpp"
#include "tensor_adaptor.hpp"
#include "threadwise_tensor_slice_transfer_v2.hpp"
#include "threadwise_contraction_dlops.hpp"
#include "threadwise_tensor_slice_transfer_v4r1.hpp"
#include "threadwise_contraction_dl.hpp"
namespace ck {
......@@ -41,7 +39,7 @@ template <index_t BlockSize,
typename enable_if<ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
bool>::type = false>
struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2
struct BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2
{
using AIndex = MultiIndex<3>;
using BIndex = MultiIndex<3>;
......@@ -148,7 +146,7 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B
MakeBBlockDescriptor_BK0_BN0_BN1_BK1(BBlockDesc_BK0_BN_BK1{});
public:
__device__ BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2()
__device__ BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2()
: c_thread_origin_data_idx_{CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
get_thread_local_1d_id())},
a_thread_copy_{
......@@ -175,6 +173,7 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B
"wrong!");
// TODO: remove this restriction
static_assert(BM0 == 2, "wrong");
static_assert(BM0 == 2 && BN0 == 2, "wrong");
}
......@@ -226,7 +225,7 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B
b_thread_desc_bk0_bn0_bn1_bk1_.GetElementSpaceSize());
constexpr auto threadwise_contraction =
ThreadwiseContractionDlops_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1<
ThreadwiseContractionDl_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1<
FloatA,
FloatB,
FloatC,
......@@ -407,4 +406,3 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B
};
} // namespace ck
#endif
include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
View file @ f9c478e2
#ifndef CK_BLOCKWISE_GEMM_XDLOPS_HPP
#define CK_BLOCKWISE_GEMM_XDLOPS_HPP
#pragma once
#include "common_header.hpp"
#include "threadwise_tensor_slice_transfer.hpp"
#include "xdlops_gemm.hpp"
#include "tensor_adaptor.hpp"
#include "thread_group.hpp"
namespace ck {
enum struct LoopScheduler
{
Default,
Interwave,
};
constexpr LoopScheduler make_default_loop_scheduler()
{
#if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
return LoopScheduler::Interwave;
#else
return LoopScheduler::Default;
#endif // if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
}
template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
......@@ -25,7 +39,9 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr index_t WaveSize = 64;
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
static constexpr index_t WaveSize = get_warp_size();
static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
......@@ -55,7 +71,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
__device__ static auto GetWaveIdx()
{
const index_t thread_id = get_thread_local_1d_id();
const index_t thread_id = ThisThreadBlock::GetThreadId();
constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
......@@ -122,8 +138,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
BK0NK1BlockDesc::IsKnownAtCompileTime(),
"wrong! Desc should be known at compile-time");
static_assert(BlockSize == MWaves * NWaves * WaveSize,
"BlockSize != MWaves * NWaves * WaveSize\n");
static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
"ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
"wrong!");
......@@ -301,7 +317,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
});
}
private:
protected:
// A[M0, M1, M2, KPerThread]
static constexpr auto a_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
......@@ -338,5 +354,232 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
};
// Note: To facilitate the inter-wave loop scheduler, we need to explicitly set the macro
// CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING=1 as a few intrinsics are not yet available in
// the latest ROCm release. For unsupported compilers, inter-wave loop scheduler falls back to the
// default loop scheduler which is given by the macro CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING=0
template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename AK0MK1BlockDesc,
typename BK0NK1BlockDesc,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack,
index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS>
struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
: public BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatAcc,
AK0MK1BlockDesc,
BK0NK1BlockDesc,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
{
using Base = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatAcc,
AK0MK1BlockDesc,
BK0NK1BlockDesc,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>;
#if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
using Base::a_block_desc_m0_m1_m2_k;
using Base::A_K1;
using Base::b_block_desc_n0_n1_n2_k;
using Base::B_K1;
using Base::c_thread_buf_;
using Base::c_thread_desc_;
using Base::CalculateAThreadOriginDataIndex;
using Base::CalculateBThreadOriginDataIndex;
using Base::I0;
using Base::I1;
using Base::KPerThread;
using Base::xdlops_gemm;
static constexpr index_t KPerInnerLoop = math::max(KPerThread / NumMacClusters, KPack);
// 2-wave optimized blockwise gemm
template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
__device__ void Run(const ABlockBuffer& a_block_buf,
const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
b_thread_desc_.GetElementSpaceSize());
static_for<0, KPerThread, KPerInnerLoop>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
// read A
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, k),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, I0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
// read B
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, k),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, I0, I0),
b_thread_buf);
});
__builtin_amdgcn_sched_barrier();
// NOTE: Synchronize threads in a workgroup at the start of each MAC cluster, but except
// the first, as we can shorten non-MAC cluster a bit and there's no observable negative
// impact. The desired effect is waves in a workgroup executing MAC in sync. This avoids
// some out-of-sync waves hijacking MAC resource from other workgroups and reducing the
// chance of latency hiding by waiting for the rest of the workgroup at the eventual
// sync point.
if constexpr(k.value != 0 || KPerInnerLoop == KPerThread)
{
asm volatile("s_barrier" ::);
__builtin_amdgcn_sched_barrier();
}
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<FloatAB, KPack> a_thread_vec;
vector_type<FloatAB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatAB>()(i) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, 0, 0, k_ + i))>{}];
b_thread_vec.template AsType<FloatAB>()(i) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, 0, 0, k_ + i))>{}];
});
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// The block_sync_lds() here performs double duty:
// A) safeguard against data hazard because barrier from blockwise_gemm is
// moved here B) reduce VMEM FIFO congestion by applying small delays to
// different wavefronts It is performed near the end of MAC cluster to
// minimize lgkmcnt penalty
if constexpr(k.value == KPerThread - KPerInnerLoop &&
k_.value == KPerInnerLoop - KPack && m0.value == MRepeat - 1 &&
n0.value == NRepeat - 1)
{
__builtin_amdgcn_sched_barrier();
block_sync_lds();
__builtin_amdgcn_sched_barrier();
}
// TODO: insert setprio in more precise manner since we
// could have more than >1 MFMA instructions in single call
xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
{
__builtin_amdgcn_sched_barrier();
__builtin_amdgcn_s_setprio(1);
__builtin_amdgcn_sched_barrier();
}
});
});
});
__builtin_amdgcn_sched_barrier();
__builtin_amdgcn_s_setprio(0);
__builtin_amdgcn_sched_barrier();
});
}
protected:
// A[M0, M1, M2, KPerInnerLoop]
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, I1, I1, Number<KPerInnerLoop>{}));
// B[N0, N1, N2, KPerInnerLoop]
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<NRepeat>{}, I1, I1, Number<KPerInnerLoop>{}));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_),
Sequence<1, 1, 1, KPerInnerLoop>,
Sequence<0, 1, 2, 3>,
3,
A_K1,
A_K1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_),
Sequence<1, 1, 1, KPerInnerLoop>,
Sequence<0, 1, 2, 3>,
3,
B_K1,
B_K1>;
AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
#endif // #if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
};
template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename AK0MK1BlockDesc,
typename BK0NK1BlockDesc,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack,
LoopScheduler LoopSched>
constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
{
if constexpr(LoopSched == LoopScheduler::Default)
{
return BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatAcc,
AK0MK1BlockDesc,
BK0NK1BlockDesc,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
}
else if constexpr(LoopSched == LoopScheduler::Interwave)
{
return BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatAcc,
AK0MK1BlockDesc,
BK0NK1BlockDesc,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
}
};
} // namespace ck
#endif
include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
View file @ f9c478e2
......@@ -45,8 +45,8 @@ struct BlockwiseTensorSliceTransfer_v5r1
src_desc, make_zero_multi_index<nDim>(), dst_desc, make_zero_multi_index<nDim>())
{
static_assert(nDim == remove_reference_t<remove_cv_t<SrcDesc>>::GetNumOfDimension() &&
nDim == remove_reference_t<remove_cv_t<DstDesc>>::GetNumOfDimension() &&
static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
nDim == BlockSliceLengths::Size() && nDim == ThreadSliceLengths::Size() &&
nDim == ThreadClusterLengths::Size() &&
nDim == ThreadClusterArrangeOrder::Size() &&
......@@ -75,14 +75,13 @@ struct BlockwiseTensorSliceTransfer_v5r1
}
}
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)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.RunRead(src_desc, src_buf, src_step_hacks);
threadwise_transfer_.RunRead(src_desc, src_buf);
}
}
......
include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v4r1.hpp → include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
View file @ f9c478e2
#ifndef CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V4R1_HPP
#define CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V4R1_HPP
#pragma once
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
......@@ -13,7 +11,7 @@ namespace ck {
// 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,
template <typename ThreadGroup,
typename SrcElementwiseOperation,
typename DstElementwiseOperation,
InMemoryDataOperationEnum DstInMemOp,
......@@ -35,7 +33,7 @@ template <index_t BlockSize,
bool ThreadTransferSrcResetCoordinateAfterRun,
bool ThreadTransferDstResetCoordinateAfterRun,
index_t NumThreadScratch = 1>
struct BlockwiseTensorSliceTransfer_v4r1
struct ThreadGroupTensorSliceTransfer_v4r1
{
static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
......@@ -43,7 +41,7 @@ struct BlockwiseTensorSliceTransfer_v4r1
using Index = MultiIndex<nDim>;
__device__ constexpr BlockwiseTensorSliceTransfer_v4r1(
__device__ constexpr ThreadGroupTensorSliceTransfer_v4r1(
const SrcDesc& src_desc,
const Index& src_block_slice_origin,
const SrcElementwiseOperation& src_element_op,
......@@ -58,8 +56,8 @@ struct BlockwiseTensorSliceTransfer_v4r1
dst_element_op)
{
static_assert(nDim == remove_reference_t<remove_cv_t<SrcDesc>>::GetNumOfDimension() &&
nDim == remove_reference_t<remove_cv_t<DstDesc>>::GetNumOfDimension() &&
static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
nDim == ThreadClusterLengths::Size() &&
nDim == ThreadClusterArrangeOrder::Size() &&
nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
......@@ -69,14 +67,14 @@ struct BlockwiseTensorSliceTransfer_v4r1
is_same<BlockSliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
"wrong! threads should be mapped to cover entire slicing window");
static_assert(BlockSize >= thread_cluster_desc_.GetElementSize(),
"wrong! BlockSize too small");
static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
"wrong! ThreadGroup::GetNumOfThread() too small");
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
make_multi_index(get_thread_local_1d_id()));
make_multi_index(ThreadGroup::GetThreadId()));
const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
......@@ -92,8 +90,8 @@ struct BlockwiseTensorSliceTransfer_v4r1
const SrcBuffer& src_buf,
Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.RunRead(src_desc, src_buf, thread_scratch_id);
}
......@@ -104,8 +102,8 @@ struct BlockwiseTensorSliceTransfer_v4r1
DstBuffer& dst_buf,
Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.RunWrite(dst_desc, dst_buf, thread_scratch_id);
}
......@@ -124,8 +122,8 @@ struct BlockwiseTensorSliceTransfer_v4r1
__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())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrcSliceWindow(src_desc, step);
}
......@@ -133,8 +131,8 @@ struct BlockwiseTensorSliceTransfer_v4r1
__device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
}
......@@ -169,4 +167,3 @@ struct BlockwiseTensorSliceTransfer_v4r1
};
} // namespace ck
#endif
include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v6r1.hpp → include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp
View file @ f9c478e2
#ifndef CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V6R1_HPP
#define CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V6R1_HPP
#pragma once
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
......@@ -13,10 +11,10 @@ namespace ck {
// 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,
template <typename ThreadGroup,
typename ElementwiseOperation,
InMemoryDataOperationEnum DstInMemOp,
typename BlockSliceLengths,
typename SliceLengths,
typename ThreadClusterLengths,
typename ThreadClusterArrangeOrder,
typename SrcData,
......@@ -28,19 +26,19 @@ template <index_t BlockSize,
index_t ScalarPerVector,
bool ThreadTransferSrcResetCoordinateAfterRun,
bool ThreadTransferDstResetCoordinateAfterRun>
struct BlockwiseTensorSliceTransfer_v6r1
struct ThreadGroupTensorSliceTransfer_v6r1
{
static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
static constexpr auto thread_slice_lengths = BlockSliceLengths{} / ThreadClusterLengths{};
static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
using Index = MultiIndex<nDim>;
__device__ constexpr BlockwiseTensorSliceTransfer_v6r1(const SrcDesc& src_desc,
const Index& src_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin,
const ElementwiseOperation& element_op)
__device__ constexpr ThreadGroupTensorSliceTransfer_v6r1(const SrcDesc& src_desc,
const Index& src_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin,
const ElementwiseOperation& element_op)
: threadwise_transfer_(src_desc,
make_zero_multi_index<nDim>(),
dst_desc,
......@@ -48,25 +46,25 @@ struct BlockwiseTensorSliceTransfer_v6r1
element_op)
{
static_assert(nDim == remove_reference_t<remove_cv_t<SrcDesc>>::GetNumOfDimension() &&
nDim == remove_reference_t<remove_cv_t<DstDesc>>::GetNumOfDimension() &&
static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
nDim == ThreadClusterLengths::Size() &&
nDim == ThreadClusterArrangeOrder::Size() &&
nDim == DimAccessOrder::Size(),
"wrong! nDim not consistent");
static_assert(
is_same<BlockSliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
"wrong! threads should be mapped to cover entire slicing window");
static_assert(BlockSize >= thread_cluster_desc_.GetElementSize(),
"wrong! BlockSize too small");
static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
"wrong! ThreadGroup::GetNumOfThread() too small");
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
make_multi_index(get_thread_local_1d_id()));
make_multi_index(ThreadGroup::GetThreadId()));
const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
......@@ -83,8 +81,8 @@ struct BlockwiseTensorSliceTransfer_v6r1
const DstDesc& dst_desc,
DstBuffer& dst_buf)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.Run(src_desc, src_buf, dst_desc, dst_buf);
}
......@@ -92,8 +90,8 @@ struct BlockwiseTensorSliceTransfer_v6r1
__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())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrcSliceWindow(src_desc, step);
}
......@@ -101,8 +99,8 @@ struct BlockwiseTensorSliceTransfer_v6r1
__device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
}
......@@ -130,4 +128,3 @@ struct BlockwiseTensorSliceTransfer_v6r1
};
} // namespace ck
#endif
include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v6r2.hpp → include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp
View file @ f9c478e2
#ifndef CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V6R2_HPP
#define CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V6R2_HPP
#pragma once
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
......@@ -13,10 +11,10 @@ namespace ck {
// 1. Use StaticallyIndexedArray instead of C array for thread buffer
// 2. It does not keep reference to tensor descriptor
// 3. Run() does not construct new tensor coordinate
template <index_t BlockSize,
template <typename ThreadGroup,
typename ElementwiseOperation,
InMemoryDataOperationEnum DstInMemOp,
typename BlockSliceLengths,
typename SliceLengths,
typename ThreadClusterLengths,
typename ThreadClusterArrangeOrder,
typename Src0Data,
......@@ -31,21 +29,21 @@ template <index_t BlockSize,
bool ThreadTransferSrc0ResetCoordinateAfterRun,
bool ThreadTransferSrc1ResetCoordinateAfterRun,
bool ThreadTransferDstResetCoordinateAfterRun>
struct BlockwiseTensorSliceTransfer_v6r2
struct ThreadGroupTensorSliceTransfer_v6r2
{
static constexpr index_t nDim = remove_reference_t<Src0Desc>::GetNumOfDimension();
static constexpr auto thread_slice_lengths = BlockSliceLengths{} / ThreadClusterLengths{};
static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
using Index = MultiIndex<nDim>;
__device__ constexpr BlockwiseTensorSliceTransfer_v6r2(const Src0Desc& src0_desc,
const Index& src0_block_slice_origin,
const Src1Desc& src1_desc,
const Index& src1_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin,
const ElementwiseOperation& element_op)
__device__ constexpr ThreadGroupTensorSliceTransfer_v6r2(const Src0Desc& src0_desc,
const Index& src0_block_slice_origin,
const Src1Desc& src1_desc,
const Index& src1_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin,
const ElementwiseOperation& element_op)
: threadwise_transfer_(src0_desc,
make_zero_multi_index<nDim>(),
src1_desc,
......@@ -55,26 +53,26 @@ struct BlockwiseTensorSliceTransfer_v6r2
element_op)
{
static_assert(nDim == remove_reference_t<remove_cv_t<Src0Desc>>::GetNumOfDimension() &&
nDim == remove_reference_t<remove_cv_t<Src1Desc>>::GetNumOfDimension() &&
nDim == remove_reference_t<remove_cv_t<DstDesc>>::GetNumOfDimension() &&
static_assert(nDim == remove_cvref_t<Src0Desc>::GetNumOfDimension() &&
nDim == remove_cvref_t<Src1Desc>::GetNumOfDimension() &&
nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
nDim == ThreadClusterLengths::Size() &&
nDim == ThreadClusterArrangeOrder::Size() &&
nDim == DimAccessOrder::Size(),
"wrong! nDim not consistent");
static_assert(
is_same<BlockSliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
"wrong! threads should be mapped to cover entire slicing window");
static_assert(BlockSize >= thread_cluster_desc_.GetElementSize(),
"wrong! BlockSize too small");
static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
"wrong! ThreadGroup::GetNumOfThread() too small");
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
make_multi_index(get_thread_local_1d_id()));
make_multi_index(ThreadGroup::GetThreadId()));
const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
......@@ -95,8 +93,8 @@ struct BlockwiseTensorSliceTransfer_v6r2
const DstDesc& dst_desc,
DstBuffer& dst_buf)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.Run(src0_desc, src0_buf, src1_desc, src1_buf, dst_desc, dst_buf);
}
......@@ -104,8 +102,8 @@ struct BlockwiseTensorSliceTransfer_v6r2
__device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc, const Index& step)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrc0SliceWindow(src0_desc, step);
}
......@@ -113,8 +111,8 @@ struct BlockwiseTensorSliceTransfer_v6r2
__device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc, const Index& step)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrc1SliceWindow(src1_desc, step);
}
......@@ -122,8 +120,8 @@ struct BlockwiseTensorSliceTransfer_v6r2
__device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
}
......@@ -154,4 +152,3 @@ struct BlockwiseTensorSliceTransfer_v6r2
};
} // namespace ck
#endif
include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v6r3.hpp → include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp
View file @ f9c478e2
#ifndef CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V6R3_HPP
#define CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V6R3_HPP
#pragma once
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
......@@ -13,10 +11,10 @@ namespace ck {
// 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,
template <typename ThreadGroup,
typename ElementwiseOperation,
InMemoryDataOperationEnum DstInMemOp,
typename BlockSliceLengths,
typename SliceLengths,
typename ThreadClusterLengths,
typename ThreadClusterArrangeOrder,
typename Src0Data,
......@@ -34,23 +32,23 @@ template <index_t BlockSize,
bool ThreadTransferSrc1ResetCoordinateAfterRun,
bool ThreadTransferSrc2ResetCoordinateAfterRun,
bool ThreadTransferDstResetCoordinateAfterRun>
struct BlockwiseTensorSliceTransfer_v6r3
struct ThreadGroupTensorSliceTransfer_v6r3
{
static constexpr index_t nDim = remove_reference_t<Src0Desc>::GetNumOfDimension();
static constexpr auto thread_slice_lengths = BlockSliceLengths{} / ThreadClusterLengths{};
static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
using Index = MultiIndex<nDim>;
__device__ constexpr BlockwiseTensorSliceTransfer_v6r3(const Src0Desc& src0_desc,
const Index& src0_block_slice_origin,
const Src1Desc& src1_desc,
const Index& src1_block_slice_origin,
const Src2Desc& src2_desc,
const Index& src2_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin,
const ElementwiseOperation& element_op)
__device__ constexpr ThreadGroupTensorSliceTransfer_v6r3(const Src0Desc& src0_desc,
const Index& src0_block_slice_origin,
const Src1Desc& src1_desc,
const Index& src1_block_slice_origin,
const Src2Desc& src2_desc,
const Index& src2_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin,
const ElementwiseOperation& element_op)
: threadwise_transfer_(src0_desc,
make_zero_multi_index<nDim>(),
src1_desc,
......@@ -62,24 +60,24 @@ struct BlockwiseTensorSliceTransfer_v6r3
element_op)
{
static_assert(nDim == remove_reference_t<remove_cv_t<Src0Desc>>::GetNumOfDimension() &&
nDim == remove_reference_t<remove_cv_t<Src1Desc>>::GetNumOfDimension() &&
nDim == remove_reference_t<remove_cv_t<Src2Desc>>::GetNumOfDimension() &&
nDim == remove_reference_t<remove_cv_t<DstDesc>>::GetNumOfDimension() &&
static_assert(nDim == remove_cvref_t<Src0Desc>::GetNumOfDimension() &&
nDim == remove_cvref_t<Src1Desc>::GetNumOfDimension() &&
nDim == remove_cvref_t<Src2Desc>::GetNumOfDimension() &&
nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
nDim == ThreadClusterLengths::Size() &&
nDim == ThreadClusterArrangeOrder::Size() &&
nDim == DimAccessOrder::Size(),
"wrong! nDim not consistent");
static_assert(
is_same<BlockSliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
"wrong! threads should be mapped to cover entire slicing window");
static_assert(BlockSize >= thread_cluster_desc_.GetElementSize(),
"wrong! BlockSize too small");
static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
"wrong! ThreadGroup::GetNumOfThread() too small");
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
make_multi_index(get_thread_local_1d_id()));
......@@ -107,8 +105,8 @@ struct BlockwiseTensorSliceTransfer_v6r3
const DstDesc& dst_desc,
DstBuffer& dst_buf)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.Run(
src0_desc, src0_buf, src1_desc, src1_buf, src2_desc, src2_buf, dst_desc, dst_buf);
......@@ -117,8 +115,8 @@ struct BlockwiseTensorSliceTransfer_v6r3
__device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc, const Index& step)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrc0SliceWindow(src0_desc, step);
}
......@@ -126,8 +124,8 @@ struct BlockwiseTensorSliceTransfer_v6r3
__device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc, const Index& step)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrc1SliceWindow(src1_desc, step);
}
......@@ -135,8 +133,8 @@ struct BlockwiseTensorSliceTransfer_v6r3
__device__ void MoveSrc2SliceWindow(const Src2Desc& src2_desc, const Index& step)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrc2SliceWindow(src2_desc, step);
}
......@@ -144,8 +142,8 @@ struct BlockwiseTensorSliceTransfer_v6r3
__device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
}
......@@ -179,4 +177,3 @@ struct BlockwiseTensorSliceTransfer_v6r3
};
} // namespace ck
#endif
include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp 0 → 100644
View file @ f9c478e2
#pragma once
namespace ck {
namespace tensor_operation {
namespace device {
enum struct ConvolutionBackwardWeightSpecialization
{
Default,
Filter1x1Stride1Pad0,
Filter1x1Pad0,
OddC,
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_base.hpp
View file @ f9c478e2
#ifndef DEVICE_BASE_HPP
#define DEVICE_BASE_HPP
#pragma once
#include <string>
#include "stream_config.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
......@@ -22,7 +23,10 @@ struct BaseInvoker
BaseInvoker(const BaseInvoker&) = default;
BaseInvoker& operator=(const BaseInvoker&) = default;
virtual float Run(const BaseArgument*, int = 1) = 0;
virtual float Run(const BaseArgument*, const StreamConfig& = StreamConfig{})
{
return float{0};
}
virtual ~BaseInvoker() {}
};
......@@ -33,8 +37,10 @@ struct BaseOperator
BaseOperator(const BaseOperator&) = default;
BaseOperator& operator=(const BaseOperator&) = default;
virtual bool IsSupportedArgument(const BaseArgument*) = 0;
virtual std::string GetTypeString() const = 0;
virtual bool IsSupportedArgument(const BaseArgument*) { return false; }
virtual std::string GetTypeString() const { return ""; }
virtual size_t GetWorkSpaceSize(const BaseArgument*) const { return 0; }
virtual ~BaseOperator() {}
};
......@@ -42,4 +48,3 @@ struct BaseOperator
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
include/ck/tensor_operation/gpu/device/device_batched_gemm_reduce_xdl_cshuffle.hpp
View file @ f9c478e2
......@@ -17,11 +17,12 @@ namespace device {
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename FloatD,
typename DPtrsGlobal,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename D1ElementwiseOperation,
typename DxsInElementwiseOperation,
typename DxsOutElementwiseOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
......@@ -37,13 +38,13 @@ __global__ void
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
FloatD* __restrict__ p_d0_grid,
FloatD* __restrict__ p_d1_grid,
DPtrsGlobal p_ds_grid,
const index_t batch_count,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
const D1ElementwiseOperation d1_element_op,
const DxsInElementwiseOperation dxs_in_element_op,
const DxsOutElementwiseOperation dxs_out_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
......@@ -64,23 +65,24 @@ __global__ void
const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch_.GetCBasePtr(g_idx)));
const long_index_t d0_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch_.GetD0BasePtr(g_idx)));
const long_index_t d1_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch_.GetD1BasePtr(g_idx)));
static_for<0, p_ds_grid.Size(), 1>{}([&](auto In) {
const long_index_t d_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch_.GetDBasePtr(g_idx, In)));
p_ds_grid(In) = p_ds_grid(In) + d_batch_offset;
});
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset,
p_c_grid + c_batch_offset,
p_d0_grid + d0_batch_offset,
p_d1_grid + d1_batch_offset,
p_ds_grid,
p_shared,
a_element_op,
b_element_op,
c_element_op,
d1_element_op,
dxs_in_element_op,
dxs_out_element_op,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mperblock_nblock_nperblock,
......@@ -90,13 +92,13 @@ __global__ void
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_c_grid;
ignore = p_d0_grid;
ignore = p_d1_grid;
ignore = p_ds_grid;
ignore = batch_count;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
ignore = d1_element_op;
ignore = dxs_in_element_op;
ignore = dxs_out_element_op;
ignore = a_grid_desc_ak0_m_ak1;
ignore = b_grid_desc_bk0_n_bk1;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
......@@ -106,6 +108,9 @@ __global__ void
#endif // end of if defined (defined(__gfx908__) || defined(__gfx90a__))
}
// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
// version currently has compiler issues with register spill which further causes validation
// failures.
template <typename ALayout,
typename BLayout,
typename CLayout,
......@@ -115,13 +120,14 @@ template <typename ALayout,
typename GemmAccDataType,
typename CShuffleDataType,
typename ReduceAccDataType,
typename DDataType,
typename DPtrsGlobal,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename D0ReduceOperation,
typename D1ReduceOperation,
typename D1ElementwiseOperation,
typename DxsReduceOperation,
typename DxsInElementwiseOperation,
typename DxsOutElementwiseOperation,
typename DGlobalMemoryDataOperation,
GemmSpecialization GemmSpec,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
......@@ -154,11 +160,14 @@ template <typename ALayout,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock>
struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwiseOperation,
index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
D1ElementwiseOperation>
DxsInElementwiseOperation,
DxsOutElementwiseOperation>
{
using DeviceOp = DeviceBatchedGemmReduce_Xdl_CShuffle;
......@@ -461,56 +470,16 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
using DGridDesc_M = decltype(MakeDGridDescriptor_M(1));
static constexpr auto MakeBlock2CTileMap(index_t batch_count,
const CGridDesc_M_N& c_grid_desc_m_n,
index_t M01,
index_t N01)
{
const auto M = c_grid_desc_m_n.GetLength(I0);
const auto N = c_grid_desc_m_n.GetLength(I1);
constexpr auto M1 = Number<MPerBlock>{};
constexpr auto N1 = Number<NPerBlock>{};
const auto M0 = M / M1;
const auto N0 = N / N1;
const auto M00 = M0 / M01;
const auto N00 = N0 / N01;
const auto g_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_insert_transform(batch_count),
make_unmerge_transform(make_tuple(M00, M01)),
make_unmerge_transform(make_tuple(N00, N01))),
make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1, 3>{}, Sequence<2, 4>{}));
const auto globalblockid_to_m00_m01_n00_n01_block_cluster_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(batch_count, M00, N00, M01, N01))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto globalblockid_to_m0_n0_block_cluster_adaptor =
chain_tensor_adaptors(g_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
globalblockid_to_m00_m01_n00_n01_block_cluster_adaptor);
return globalblockid_to_m0_n0_block_cluster_adaptor;
}
struct ComputeBasePtrOfStridedBatch
{
ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
index_t BatchStrideB,
index_t BatchStrideC,
index_t BatchStrideD0,
index_t BatchStrideD1)
index_t BatchStrideD)
: BatchStrideA_(BatchStrideA),
BatchStrideB_(BatchStrideB),
BatchStrideC_(BatchStrideC),
BatchStrideD0_(BatchStrideD0),
BatchStrideD1_(BatchStrideD1)
BatchStrideD_(BatchStrideD)
{
}
......@@ -529,22 +498,20 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
return g_idx * static_cast<long_index_t>(BatchStrideC_);
}
__host__ __device__ constexpr long_index_t GetD0BasePtr(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideD0_);
}
__host__ __device__ constexpr long_index_t GetD1BasePtr(index_t g_idx) const
template <index_t I>
__host__ __device__ constexpr long_index_t GetDBasePtr(index_t g_idx,
Number<I> reduction_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideD1_);
// TODO - Support sequence of StrideD in MakeArgument()
(void)reduction_idx;
return g_idx * static_cast<long_index_t>(BatchStrideD_);
}
private:
index_t BatchStrideA_;
index_t BatchStrideB_;
index_t BatchStrideC_;
index_t BatchStrideD0_;
index_t BatchStrideD1_;
index_t BatchStrideD_;
};
// GridwiseGemm
......@@ -554,15 +521,15 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
CShuffleDataType,
CDataType,
ReduceAccDataType,
DDataType,
DPtrsGlobal,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
D0ReduceOperation,
D1ReduceOperation,
D1ElementwiseOperation,
DxsReduceOperation,
DxsInElementwiseOperation,
DxsOutElementwiseOperation,
InMemoryDataOperationEnum::Set,
InMemoryDataOperationEnum::AtomicAdd,
DGlobalMemoryDataOperation,
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
CGridDesc_M_N,
......@@ -600,9 +567,8 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
CShuffleBlockTransferScalarPerVector_NPerBlock,
CReduceThreadClusterLengths_MPerBlock_NPerBlock,
CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock>;
using Block2CTileMap = decltype(MakeBlock2CTileMap(1, CGridDesc_M_N{}, 1, 1));
CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
LoopSched>;
// Argument
struct Argument : public BaseArgument
......@@ -610,8 +576,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
Argument(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
DDataType* p_d0_grid,
DDataType* p_d1_grid,
DPtrsGlobal p_ds_grid,
index_t MRaw,
index_t NRaw,
index_t KRaw,
......@@ -621,13 +586,13 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
D1ElementwiseOperation d1_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsOutElementwiseOperation dxs_out_element_op,
index_t BatchCount)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_c_grid_{p_c_grid},
p_d0_grid_{p_d0_grid},
p_d1_grid_{p_d1_grid},
p_ds_grid_{p_ds_grid},
BatchCount_(BatchCount),
a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
......@@ -635,19 +600,22 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
d_grid_desc_m_{DeviceOp::MakeDGridDescriptor_M(MRaw)},
c_grid_desc_mblock_mperblock_nblock_nperblock_{},
d_grid_desc_mblock_mperblock_{},
compute_base_ptr_of_batch_{a_grid_desc_ak0_m_ak1_.GetElementSpaceSize(),
b_grid_desc_bk0_n_bk1_.GetElementSpaceSize(),
c_grid_desc_m_n_.GetElementSpaceSize(),
d_grid_desc_m_.GetElementSpaceSize(),
d_grid_desc_m_.GetElementSpaceSize()},
block_2_ctile_map_{},
compute_base_ptr_of_batch_{
type_convert<index_t>(a_grid_desc_ak0_m_ak1_.GetElementSpaceSize()),
type_convert<index_t>(b_grid_desc_bk0_n_bk1_.GetElementSpaceSize()),
type_convert<index_t>(c_grid_desc_m_n_.GetElementSpaceSize()),
type_convert<index_t>(d_grid_desc_m_.GetElementSpaceSize())},
block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
c_element_op_{c_element_op},
d1_element_op_{d1_element_op}
dxs_in_element_op_{dxs_in_element_op},
dxs_out_element_op_{dxs_out_element_op}
{
if(GridwiseGemm::CheckValidity(
a_grid_desc_ak0_m_ak1_, b_grid_desc_bk0_n_bk1_, c_grid_desc_m_n_))
if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
b_grid_desc_bk0_n_bk1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
......@@ -655,8 +623,6 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
d_grid_desc_mblock_mperblock_ =
GridwiseGemm::MakeDGridDescriptor_MBlock_MPerBlock(d_grid_desc_m_);
block_2_ctile_map_ = MakeBlock2CTileMap(BatchCount, c_grid_desc_m_n_, 1, 1);
}
}
......@@ -664,8 +630,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
DDataType* p_d0_grid_;
DDataType* p_d1_grid_;
DPtrsGlobal p_ds_grid_;
index_t BatchCount_;
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
......@@ -675,11 +640,12 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
c_grid_desc_mblock_mperblock_nblock_nperblock_;
typename GridwiseGemm::DGridDescriptor_MBlock_MPerBlock d_grid_desc_mblock_mperblock_;
ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
Block2CTileMap block_2_ctile_map_;
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
D1ElementwiseOperation d1_element_op_;
DxsInElementwiseOperation dxs_in_element_op_;
DxsOutElementwiseOperation dxs_out_element_op_;
};
// Invoker
......@@ -687,7 +653,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, int /* nrepeat */ = 1)
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
#if 0
{
......@@ -711,58 +677,63 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
}
#endif
if(!GridwiseGemm::CheckValidity(
arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.c_grid_desc_m_n_))
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_))
{
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
const index_t grid_size =
GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_) * arg.BatchCount_;
const auto K0 = arg.a_grid_desc_ak0_m_ak1_.GetLength(I0);
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.BatchCount_;
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
if(has_main_k0_block_loop)
float elapsed_time = 0.0f;
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_batched_gemm_reduce_xdl_cshuffle_v1<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
DDataType,
DPtrsGlobal,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
D1ElementwiseOperation,
DxsInElementwiseOperation,
DxsOutElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename GridwiseGemm::DGridDescriptor_MBlock_MPerBlock,
ComputeBasePtrOfStridedBatch,
remove_reference_t<Block2CTileMap>,
typename GridwiseGemm::DefaultBlock2CTileMap,
true>;
launch_kernel(kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.p_d0_grid_,
arg.p_d1_grid_,
arg.BatchCount_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.d1_element_op_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.d_grid_desc_mblock_mperblock_,
arg.compute_base_ptr_of_batch_,
arg.block_2_ctile_map_);
elapsed_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.p_ds_grid_,
arg.BatchCount_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.dxs_in_element_op_,
arg.dxs_out_element_op_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.d_grid_desc_mblock_mperblock_,
arg.compute_base_ptr_of_batch_,
arg.block_2_ctile_map_);
}
else
{
......@@ -770,48 +741,52 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
DDataType,
DPtrsGlobal,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
D1ElementwiseOperation,
DxsInElementwiseOperation,
DxsOutElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename GridwiseGemm::DGridDescriptor_MBlock_MPerBlock,
ComputeBasePtrOfStridedBatch,
remove_reference_t<Block2CTileMap>,
typename GridwiseGemm::DefaultBlock2CTileMap,
false>;
launch_kernel(kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.p_d0_grid_,
arg.p_d1_grid_,
arg.BatchCount_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.d1_element_op_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.d_grid_desc_mblock_mperblock_,
arg.compute_base_ptr_of_batch_,
arg.block_2_ctile_map_);
elapsed_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.p_ds_grid_,
arg.BatchCount_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.dxs_in_element_op_,
arg.dxs_out_element_op_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.d_grid_desc_mblock_mperblock_,
arg.compute_base_ptr_of_batch_,
arg.block_2_ctile_map_);
}
return 0;
return elapsed_time;
}
// polymorphic
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
......@@ -823,8 +798,10 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
static bool IsSupportedArgument(const Argument& arg)
{
return GridwiseGemm::CheckValidity(
arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.c_grid_desc_m_n_);
return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_);
}
// polymorphic
......@@ -844,8 +821,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
CDataType* p_c,
DDataType* p_d0,
DDataType* p_d1,
DPtrsGlobal p_dxs,
index_t MRaw,
index_t NRaw,
index_t KRaw,
......@@ -855,14 +831,14 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
D1ElementwiseOperation d1_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsOutElementwiseOperation dxs_out_element_op,
index_t BatchCount)
{
return Argument{p_a,
p_b,
p_c,
p_d0,
p_d1,
p_dxs,
MRaw,
NRaw,
KRaw,
......@@ -872,7 +848,8 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
a_element_op,
b_element_op,
c_element_op,
d1_element_op,
dxs_in_element_op,
dxs_out_element_op,
BatchCount};
}
......@@ -882,8 +859,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
void* p_d0,
void* p_d1,
DPtrsGlobal p_dxs,
index_t MRaw,
index_t NRaw,
index_t KRaw,
......@@ -893,14 +869,14 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
D1ElementwiseOperation d1_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsOutElementwiseOperation dxs_out_element_op,
index_t BatchCount) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
static_cast<DDataType*>(p_d0),
static_cast<DDataType*>(p_d1),
p_dxs,
MRaw,
NRaw,
KRaw,
......@@ -910,7 +886,8 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
a_element_op,
b_element_op,
c_element_op,
d1_element_op,
dxs_in_element_op,
dxs_out_element_op,
BatchCount);
}
......
include/ck/tensor_operation/gpu/device/device_batched_gemm_xdl.hpp
View file @ f9c478e2
......@@ -243,44 +243,6 @@ struct DeviceBatchedGemmXdl
using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
static constexpr auto MakeBlock2CTileMap(index_t batch_count,
const CGridDesc_M_N& c_grid_desc_m_n,
index_t M01,
index_t N01)
{
const auto M = c_grid_desc_m_n.GetLength(I0);
const auto N = c_grid_desc_m_n.GetLength(I1);
constexpr auto M1 = Number<MPerBlock>{};
constexpr auto N1 = Number<NPerBlock>{};
const auto M0 = M / M1;
const auto N0 = N / N1;
const auto M00 = M0 / M01;
const auto N00 = N0 / N01;
const auto g_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_insert_transform(batch_count),
make_unmerge_transform(make_tuple(M00, M01)),
make_unmerge_transform(make_tuple(N00, N01))),
make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1, 3>{}, Sequence<2, 4>{}));
const auto globalblockid_to_m00_m01_n00_n01_block_cluster_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(batch_count, M00, N00, M01, N01))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto globalblockid_to_m0_n0_block_cluster_adaptor =
chain_tensor_adaptors(g_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
globalblockid_to_m00_m01_n00_n01_block_cluster_adaptor);
return globalblockid_to_m0_n0_block_cluster_adaptor;
}
struct ComputePtrOffsetOfStridedBatch
{
ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
......@@ -354,7 +316,7 @@ struct DeviceBatchedGemmXdl
using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
using Block2CTileMap = decltype(MakeBlock2CTileMap(1, CGridDesc_M_N{}, 1, 1));
using Block2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
// Argument
struct Argument : public BaseArgument
......@@ -384,23 +346,25 @@ struct DeviceBatchedGemmXdl
DeviceBatchedGemmXdl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB)},
c_grid_desc_m_n_{DeviceBatchedGemmXdl::MakeCGridDescriptor_M_N(M, N, StrideC)},
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
compute_ptr_offset_of_batch_{a_grid_desc_k0_m_k1_.GetElementSpaceSize(),
b_grid_desc_k0_n_k1_.GetElementSpaceSize(),
c_grid_desc_m_n_.GetElementSpaceSize()},
block_2_ctile_map_{},
compute_ptr_offset_of_batch_{
type_convert<index_t>(a_grid_desc_k0_m_k1_.GetElementSpaceSize()),
type_convert<index_t>(b_grid_desc_k0_n_k1_.GetElementSpaceSize()),
type_convert<index_t>(c_grid_desc_m_n_.GetElementSpaceSize())},
block_2_ctile_map_{
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01)},
M01_{M01},
N01_{N01},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
c_element_op_{c_element_op}
{
if(GridwiseGemm::CheckValidity(
a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
block_2_ctile_map_ = MakeBlock2CTileMap(BatchCount, c_grid_desc_m_n_, M01, N01);
}
}
......@@ -427,7 +391,7 @@ struct DeviceBatchedGemmXdl
{
using Argument = DeviceBatchedGemmXdl::Argument;
float Run(const Argument& arg, int nrepeat = 1)
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
{
std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
......@@ -445,23 +409,21 @@ struct DeviceBatchedGemmXdl
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
arg.block_2_ctile_map_))
{
throw std::runtime_error(
"wrong! GridwiseBatchedGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting");
}
const index_t grid_size =
GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_) * arg.BatchCount_;
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.BatchCount_;
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_batched_gemm_xdlops_v2r3<
GridwiseGemm,
......@@ -477,8 +439,8 @@ struct DeviceBatchedGemmXdl
remove_reference_t<Block2CTileMap>,
true>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -511,8 +473,8 @@ struct DeviceBatchedGemmXdl
remove_reference_t<Block2CTileMap>,
false>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -534,9 +496,10 @@ struct DeviceBatchedGemmXdl
}
// polymorphic
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
......@@ -551,8 +514,7 @@ struct DeviceBatchedGemmXdl
return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_);
arg.block_2_ctile_map_);
}
// polymorphic
......
include/ck/tensor_operation/gpu/device/device_binary_elementwise.hpp 0 → 100644
View file @ f9c478e2
#pragma once
#include <iostream>
#include <vector>
#include "device.hpp"
#include "device_base.hpp"
#include "gridwise_binary_elementwise_1d.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ComputeDataType,
typename ElementwiseFunctor,
index_t NDim,
index_t MPerThread,
index_t AScalarPerVector,
index_t BScalarPerVector,
index_t CScalarPerVector>
struct DeviceBinaryElementwise : public BaseOperator
{
static constexpr auto I0 = Number<0>{};
template <typename Desc_M>
static auto PadDescriptor_M_1d(Desc_M desc_m, index_t gridSize, index_t blockSize)
{
const auto M = desc_m.GetLength(I0);
const index_t loop_step = gridSize * blockSize * MPerThread;
const auto pad = math::integer_least_multiple(M, loop_step) - M;
const auto desc_m_pad =
transform_tensor_descriptor(desc_m,
make_tuple(make_right_pad_transform(M, pad)),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0>{}));
return desc_m_pad;
}
static auto MakeDescriptor_M(const std::vector<index_t>& lengths,
const std::vector<index_t>& strides,
index_t gridSize,
index_t blockSize)
{
auto tupleOfShape = generate_tuple([&](auto I) { return lengths[I]; }, Number<NDim>{});
auto tupleOfStride = generate_tuple([&](auto I) { return strides[I]; }, Number<NDim>{});
// nd desc - [s0, s1, s2, ...]
const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
// merge nd to 1d desc - [s0 * s1 * ...]
if constexpr(NDim > 1)
{
const auto desc_m = transform_tensor_descriptor(
desc,
make_tuple(make_merge_transform(tupleOfShape)),
make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<NDim>{})),
make_tuple(Sequence<0>{}));
return PadDescriptor_M_1d(desc_m, gridSize, blockSize);
}
else
return PadDescriptor_M_1d(desc, gridSize, blockSize);
}
using AGridDesc_M = decltype(MakeDescriptor_M({1, 1}, {1, 1}, 1, 1));
using BGridDesc_M = decltype(MakeDescriptor_M({1, 1}, {1, 1}, 1, 1));
using CGridDesc_M = decltype(MakeDescriptor_M({1, 1}, {1, 1}, 1, 1));
using GridwiseBinEltwise = GridwiseBinaryElementwise_1D<ADataType,
BDataType,
CDataType,
ComputeDataType,
AGridDesc_M,
BGridDesc_M,
CGridDesc_M,
ElementwiseFunctor,
MPerThread,
AScalarPerVector,
BScalarPerVector,
CScalarPerVector>;
struct Argument : public BaseArgument
{
Argument(const ADataType* p_a,
const BDataType* p_b,
CDataType* p_c,
const std::vector<index_t>& lengths,
const std::vector<index_t>& a_strides,
const std::vector<index_t>& b_strides,
const std::vector<index_t>& c_strides,
ElementwiseFunctor functor)
: p_a_(p_a),
p_b_(p_b),
p_c_(p_c),
lengths_(lengths),
a_strides_(a_strides),
b_strides_(b_strides),
c_strides_(c_strides),
functor_(functor),
blockSize_(256),
gridSize_(120) // FIXME - Calculate the grid size by number of CU in the future
{
a_grid_desc_m_ = MakeDescriptor_M(lengths, a_strides, gridSize_, blockSize_);
b_grid_desc_m_ = MakeDescriptor_M(lengths, b_strides, gridSize_, blockSize_);
c_grid_desc_m_ = MakeDescriptor_M(lengths, c_strides, gridSize_, blockSize_);
}
const ADataType* p_a_;
const BDataType* p_b_;
CDataType* p_c_;
std::vector<int> lengths_;
AGridDesc_M a_grid_desc_m_;
BGridDesc_M b_grid_desc_m_;
CGridDesc_M c_grid_desc_m_;
std::vector<index_t> a_strides_;
std::vector<index_t> b_strides_;
std::vector<index_t> c_strides_;
ElementwiseFunctor functor_;
index_t blockSize_;
index_t gridSize_;
};
struct Invoker : public BaseInvoker
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
const auto kernel = kernel_binary_elementwise_1d<GridwiseBinEltwise,
ADataType,
BDataType,
CDataType,
AGridDesc_M,
BGridDesc_M,
CGridDesc_M,
ElementwiseFunctor>;
float elapsed_time = launch_and_time_kernel(stream_config,
kernel,
dim3(arg.gridSize_),
dim3(arg.blockSize_),
0,
arg.p_a_,
arg.p_b_,
arg.p_c_,
arg.a_grid_desc_m_,
arg.b_grid_desc_m_,
arg.c_grid_desc_m_,
arg.functor_);
return elapsed_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
if(pArg == nullptr)
return false;
if(pArg->lengths_.size() != NDim)
return false;
if(pArg->lengths_.back() % MPerThread != 0)
return false;
auto IsScalarPerVectorValid = [](bool isLastDimensionCoalesced, int scalarPerVector) {
bool ret = true;
if(!isLastDimensionCoalesced)
ret = scalarPerVector == 1;
else
ret = MPerThread % scalarPerVector == 0;
return ret;
};
if(!IsScalarPerVectorValid(pArg->a_strides_.back() == 1, AScalarPerVector))
return false;
if(!IsScalarPerVectorValid(pArg->b_strides_.back() == 1, BScalarPerVector))
return false;
if(!IsScalarPerVectorValid(pArg->c_strides_.back() == 1, CScalarPerVector))
return false;
return true;
};
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
std::vector<index_t> lengths,
std::vector<index_t> a_strides,
std::vector<index_t> b_strides,
std::vector<index_t> c_strides,
ElementwiseFunctor functor)
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
lengths,
a_strides,
b_strides,
c_strides,
functor);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() { return std::make_unique<Invoker>(); }
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceBinaryElementwise"
<< "<"
<< "MPerThread = " << MPerThread
<< ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
View file @ f9c478e2
......@@ -11,7 +11,7 @@
#include "tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_xdlops_v2r4r2.hpp"
#include "gridwise_gemm_xdlops_bwd_weight.hpp"
namespace ck {
namespace tensor_operation {
......@@ -81,6 +81,22 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
static constexpr auto K1Number = Number<K1>{};
static constexpr auto GemmK1Number = K1Number;
static constexpr auto N1Number = K1Number;
// Bytes per 32 lds bank: 32 * 4 bytes
static constexpr auto BankLength = 128;
static constexpr auto ElePerBank = BankLength / sizeof(ADataType);
// M1 & M0
static constexpr auto ABlockLdsM1PerBlock = ElePerBank / K1;
static constexpr auto ABlockLdsM0PerBlock = MPerBlock / ABlockLdsM1PerBlock;
static constexpr auto ABlockLdsM1Padding = 4;
// N1 & N0
static constexpr auto BBlockLdsN1PerBlock = ElePerBank / K1;
static constexpr auto BBlockLdsN0PerBlock = NPerBlock / BBlockLdsN1PerBlock;
static constexpr auto BBlockLdsN1Padding = 4;
static auto
MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
ck::index_t K,
......@@ -125,27 +141,51 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
const index_t GemmK0 =
math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
K0PerBlock;
const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
const auto out_gemmktotal_gemmm_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
const auto in_n_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
// A: output tensor
const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
out_gemmktotal_gemmm_grid_desc,
make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_pass_through_transform(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const index_t N0 = N / N1Number;
const index_t GemmK0Total = N0 * Ho * Wo;
const index_t GemmK0S =
math::integer_divide_ceil(GemmK0Total, K0PerBlock * GemmKBatch) * K0PerBlock;
const index_t GemmK0Pad = GemmKBatch * GemmK0S;
const auto out_n_ho_wo_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Ho * Wo, K));
const auto out_n0_ho_wo_k_n1_grid_desc =
transform_tensor_descriptor(out_n_ho_wo_k_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(N0, N1Number)),
make_pass_through_transform(Ho * Wo),
make_pass_through_transform(K)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 3>{}, Sequence<1>{}, Sequence<2>{}));
const auto out_gemmk0total_gemmm_gemmk1_grid_desc =
transform_tensor_descriptor(out_n0_ho_wo_k_n1_grid_desc,
make_tuple(make_merge_transform(make_tuple(N0, Ho * Wo)),
make_pass_through_transform(K),
make_pass_through_transform(N1Number)),
make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto out_gemmk0pad_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
out_gemmk0total_gemmm_gemmk1_grid_desc,
make_tuple(make_right_pad_transform(GemmK0Total, GemmK0Pad - GemmK0Total),
make_pass_through_transform(GemmM),
make_pass_through_transform(N1Number)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
out_gemmkpad_gemmm_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
make_pass_through_transform(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
out_gemmk0pad_gemmm_gemmk1_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0)),
make_pass_through_transform(GemmM),
make_pass_through_transform(N1Number)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}));
// B: input tensor
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
......@@ -167,26 +207,50 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
const auto in_gemmktotal_gemmn_grid_desc =
const auto in_n0_y_ho_x_wo_c_n1_grid_desc =
transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(Y, X, C)),
make_merge_transform(make_tuple(N, Ho, Wo))),
make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
in_gemmktotal_gemmn_grid_desc,
make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
make_tuple(make_unmerge_transform(make_tuple(N0, N1Number)),
make_pass_through_transform(Y),
make_pass_through_transform(Ho),
make_pass_through_transform(X),
make_pass_through_transform(Wo),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{},
Sequence<1>{},
Sequence<2>{},
Sequence<3>{},
Sequence<4>{},
Sequence<5>{}),
make_tuple(Sequence<0, 6>{},
Sequence<1>{},
Sequence<2>{},
Sequence<3>{},
Sequence<4>{},
Sequence<5>{}));
const auto in_gemmk0total_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
in_n0_y_ho_x_wo_c_n1_grid_desc,
make_tuple(make_merge_transform(make_tuple(N0, Ho, Wo)),
make_merge_transform(make_tuple(Y, X, C)),
make_pass_through_transform(N1Number)),
make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}, Sequence<6>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto in_gemmk0pad_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
in_gemmk0total_gemmn_gemmk1_grid_desc,
make_tuple(make_right_pad_transform(GemmK0Total, GemmK0Pad - GemmK0Total),
make_pass_through_transform(GemmN),
make_pass_through_transform(N1Number)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
in_gemmkpad_gemmn_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
in_gemmk0pad_gemmn_gemmk1_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0)),
make_pass_through_transform(GemmN),
make_pass_through_transform(N1Number)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc =
......@@ -205,7 +269,7 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
using CGridDesc_M_N = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
AccDataType,
......@@ -233,6 +297,9 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
ABlockTransferDstScalarPerVector_K1,
false, // AThreadTransferSrcResetCoordinateAfterRun,
ABlockLdsAddExtraM,
ABlockLdsM1PerBlock,
ABlockLdsM0PerBlock,
ABlockLdsM1Padding,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
......@@ -241,12 +308,17 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
BBlockTransferDstScalarPerVector_K1,
false, // BThreadTransferSrcResetCoordinateAfterRun,
BBlockLdsAddExtraN,
BBlockLdsN1PerBlock,
BBlockLdsN0PerBlock,
BBlockLdsN1Padding,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CBlockTransferScalarPerVector_NWaveNPerXdl,
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
true,
true>;
using GridwiseGemmAtomicAdd = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
using GridwiseGemmAtomicAdd = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
AccDataType,
......@@ -274,6 +346,9 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
ABlockTransferDstScalarPerVector_K1,
false, // AThreadTransferSrcResetCoordinateAfterRun,
ABlockLdsAddExtraM,
ABlockLdsM1PerBlock,
ABlockLdsM0PerBlock,
ABlockLdsM1Padding,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
......@@ -282,10 +357,15 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
BBlockTransferDstScalarPerVector_K1,
false, // BThreadTransferSrcResetCoordinateAfterRun,
BBlockLdsAddExtraN,
BBlockLdsN1PerBlock,
BBlockLdsN0PerBlock,
BBlockLdsN1Padding,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CBlockTransferScalarPerVector_NWaveNPerXdl,
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
true,
true>;
// Argument
using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
decltype(GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}));
......@@ -353,17 +433,16 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
b_grid_desc_kbatch_k0_n_k1_ = descs[I1];
c_grid_desc_m_n_ = descs[I2];
block_2_ctile_map_ =
GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
if(GridwiseGemm::CheckValidity(a_grid_desc_kbatch_k0_m_k1_,
b_grid_desc_kbatch_k0_n_k1_,
c_grid_desc_m_n_,
M01_,
N01_))
block_2_ctile_map_))
{
c_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n_);
block_2_ctile_map_ =
GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
}
}
......@@ -415,20 +494,21 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
}
float Run(const Argument& arg, int nrepeat = 1)
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
ShowInfo(arg);
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
arg.block_2_ctile_map_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
"wrong! GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight has invalid setting");
}
const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0);
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_, kbatch);
const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0);
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1);
......@@ -437,56 +517,35 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
float ave_time = 0;
const auto Run = [&](const auto& kernel) {
if(nrepeat > 0)
{
ave_time =
launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
}
if(kbatch > 1 || nrepeat <= 0)
{
hipGetErrorString(hipMemset(
arg.p_c_grid_,
0,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
sizeof(CDataType)));
launch_kernel(kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
}
hipGetErrorString(hipMemset(
arg.p_c_grid_,
0,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
sizeof(CDataType)));
ave_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
};
if(has_main_k0_block_loop)
{
if(kbatch == 1)
{
const auto kernel = kernel_gemm_xdlops_v2r4r2<
const auto kernel = kernel_gemm_xdlops_bwd_weight<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
......@@ -503,7 +562,7 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
}
else
{
const auto kernel = kernel_gemm_xdlops_v2r4r2<
const auto kernel = kernel_gemm_xdlops_bwd_weight<
GridwiseGemmAtomicAdd,
ADataType, // TODO: distiguish A/B datatype
CDataType,
......@@ -523,7 +582,7 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
{
if(kbatch == 1)
{
const auto kernel = kernel_gemm_xdlops_v2r4r2<
const auto kernel = kernel_gemm_xdlops_bwd_weight<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
......@@ -540,7 +599,7 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
}
else
{
const auto kernel = kernel_gemm_xdlops_v2r4r2<
const auto kernel = kernel_gemm_xdlops_bwd_weight<
GridwiseGemmAtomicAdd,
ADataType, // TODO: distiguish A/B datatype
CDataType,
......@@ -560,9 +619,10 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
return ave_time;
}
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
......@@ -582,6 +642,12 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
return false;
}
// unmerge N to N0 and N1, where N1 equals to K1
if(!(arg.Conv_N_ % K1 == 0))
{
return false;
}
// vector store C matrix into global memory
if(!(arg.Conv_C_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
{
......@@ -592,8 +658,7 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
return GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_);
arg.block_2_ctile_map_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
......
include/ck/tensor_operation/gpu/device/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
View file @ f9c478e2
......@@ -486,13 +486,16 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
c_grid_desc_m_n_container_.push_back(descs[I2]);
if(GridwiseGemm::CheckValidity(descs[I0], descs[I1], descs[I2], M01_, N01_))
auto block_2_ctile_map =
GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2], M01, N01);
if(GridwiseGemm::CheckValidity(
descs[I0], descs[I1], descs[I2], block_2_ctile_map))
{
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_.push_back(
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(descs[I2]));
block_2_ctile_map_container_.push_back(
GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2], M01, N01));
block_2_ctile_map_container_.push_back(block_2_ctile_map);
}
}
}
......@@ -531,7 +534,7 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, int nrepeat = 1)
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
float ave_time = 0;
for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
......@@ -572,21 +575,19 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
arg.b_grid_desc_k0_n_k1_container_[i],
arg.c_grid_desc_m_n_container_[i],
arg.M01_,
arg.N01_))
arg.block_2_ctile_map_container_[i]))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
}
const index_t grid_size =
GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_container_[i]);
const auto K0 = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0);
const index_t grid_size = arg.block_2_ctile_map_container_[i].CalculateGridSize(
arg.c_grid_desc_m_n_container_[i]);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) *
arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2);
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,
......@@ -603,8 +604,8 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
true>;
ave_time += launch_and_time_kernel(
stream_config,
kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -636,8 +637,8 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
false>;
ave_time += launch_and_time_kernel(
stream_config,
kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -656,9 +657,10 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
return ave_time;
}
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
......@@ -698,13 +700,12 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
}
// Gridwise GEMM size
for(int i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
for(std::size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
{
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
arg.b_grid_desc_k0_n_k1_container_[i],
arg.c_grid_desc_m_n_container_[i],
arg.M01_,
arg.N01_))
arg.block_2_ctile_map_container_[i]))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
View file @ f9c478e2
......@@ -540,7 +540,8 @@ struct
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
block_2_ctile_map_{},
block_2_ctile_map_{
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01)},
M01_{M01},
N01_{N01},
in_element_op_{in_element_op},
......@@ -575,8 +576,10 @@ struct
c0_grid_desc_m_n_ = descs[I3];
c1_grid_desc_m_n_ = descs[I4];
if(GridwiseGemm::CheckValidity(
a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
GridwiseGemm::
......@@ -592,9 +595,6 @@ struct
GridwiseGemm::
MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
c1_grid_desc_m_n_);
block_2_ctile_map_ =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
}
}
......@@ -642,7 +642,7 @@ struct
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, int nrepeat = 1)
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
#if 0
{
......@@ -689,22 +689,21 @@ struct
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
arg.block_2_ctile_map_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r3 has invalid setting");
}
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v3r3<
GridwiseGemm,
......@@ -728,8 +727,8 @@ struct
true>;
ave_time = launch_and_time_kernel(
stream_config,
kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -772,8 +771,8 @@ struct
false>;
ave_time = launch_and_time_kernel(
stream_config,
kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -796,9 +795,10 @@ struct
return ave_time;
}
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
......@@ -852,8 +852,7 @@ struct
return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_);
arg.block_2_ctile_map_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
......
include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
View file @ f9c478e2
#ifndef DEVICE_CONV2D_FWD_XDL_C_SHUFFLE_BIAS_ACTIVATION_NHWC_KYXC_NHWK_HPP
#define DEVICE_CONV2D_FWD_XDL_C_SHUFFLE_BIAS_ACTIVATION_NHWC_KYXC_NHWK_HPP
#pragma once
#include <iostream>
#include <sstream>
#include "device.hpp"
......@@ -550,9 +548,13 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
b_grid_desc_k0_n_k1_ = descs[I1];
c_grid_desc_m_n_ = descs[I2];
c0_grid_desc_m_n_ = descs[I3];
block_2_ctile_map_ =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
if(GridwiseGemm::CheckValidity(
a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
GridwiseGemm::
......@@ -563,9 +565,6 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
GridwiseGemm::
MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
c0_grid_desc_m_n_);
block_2_ctile_map_ =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
}
}
......@@ -607,7 +606,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, int nrepeat = 1)
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
#if 0
{
......@@ -651,22 +650,21 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
arg.block_2_ctile_map_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r2 has invalid setting");
}
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v3r2<
GridwiseGemm,
......@@ -687,8 +685,8 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
true>;
ave_time = launch_and_time_kernel(
stream_config,
kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -726,8 +724,8 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
false>;
ave_time = launch_and_time_kernel(
stream_config,
kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -748,9 +746,10 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
return ave_time;
}
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
......@@ -804,8 +803,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_);
arg.block_2_ctile_map_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
......@@ -919,4 +917,3 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
View file @ f9c478e2
......@@ -417,6 +417,8 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
using CGridDesc_M_N = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
using Block2CTileMap = BlockToCTileMap_M00_N0_M01<MPerBlock, NPerBlock, CGridDesc_M_N>;
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1<
BlockSize,
......@@ -477,8 +479,6 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
ck::index_t M01,
ck::index_t N01,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
......@@ -490,8 +490,6 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
c_grid_desc_m_n_{},
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
in_element_op_{in_element_op},
wei_element_op_{wei_element_op},
out_element_op_{out_element_op},
......@@ -522,16 +520,17 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
b_grid_desc_k0_n_k1_ = descs[I1];
c_grid_desc_m_n_ = descs[I2];
if(GridwiseGemm::CheckValidity(
a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
block_2_ctile_map_ = Block2CTileMap{c_grid_desc_m_n_};
if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
GridwiseGemm::
MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
c_grid_desc_m_n_);
block_2_ctile_map_ =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
}
}
......@@ -544,9 +543,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
typename GridwiseGemm::
CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
Block2CTileMap block_2_ctile_map_;
InElementwiseOperation in_element_op_;
WeiElementwiseOperation wei_element_op_;
OutElementwiseOperation out_element_op_;
......@@ -568,7 +565,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, int nrepeat = 1)
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
#if 0
{
......@@ -631,22 +628,21 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
arg.block_2_ctile_map_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
}
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v3r1<
GridwiseGemm,
......@@ -660,12 +656,12 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
Block2CTileMap,
true>;
ave_time = launch_and_time_kernel(
stream_config,
kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -694,12 +690,12 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
Block2CTileMap,
false>;
ave_time = launch_and_time_kernel(
stream_config,
kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -718,9 +714,10 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
return ave_time;
}
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
......@@ -774,8 +771,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_);
arg.block_2_ctile_map_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
......@@ -813,8 +809,6 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op};
......@@ -853,8 +847,6 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op);
......
include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
View file @ f9c478e2
......@@ -408,15 +408,16 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
a_grid_desc_k0_m_k1_ = descs[I0];
b_grid_desc_k0_n_k1_ = descs[I1];
c_grid_desc_m_n_ = descs[I2];
block_2_ctile_map_ =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
if(GridwiseGemm::CheckValidity(
a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
block_2_ctile_map_ =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
}
}
......@@ -450,7 +451,7 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, int nrepeat = 1)
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
#if 0
{
......@@ -469,22 +470,21 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
arg.block_2_ctile_map_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
}
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,
......@@ -499,8 +499,8 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
true>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -530,8 +530,8 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
false>;
ave_time = launch_and_time_kernel(kernel,
nrepeat,
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
......@@ -550,9 +550,10 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
return ave_time;
}
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
......@@ -606,8 +607,7 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_);
arg.block_2_ctile_map_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
......
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