Skip to content

GitLab

Commit e6a23d8b authored by Jing Zhang's avatar Jing Zhang
Browse files

add e2

parent a8169558
Hide whitespace changes
Inline Side-by-side
Showing with 322 additions and 287 deletions
composable_kernel/include/tensor_operation/blockwise_gemm_dlops_v3.hpp
View file @ e6a23d8b
......@@ -7,20 +7,21 @@
namespace ck {
template <index_t BlockSize,
typename FloatAB,
typename FloatA,
typename FloatB,
typename FloatC,
typename BlockMatrixA,
typename BlockMatrixB,
typename ThreadMatrixC,
typename ABlockDesc_E1_K_E2,
typename BBlockDesc_E1_N_Ho_Wo_E2,
typename CThreadDesc_K_N_Ho_Wo,
index_t EPerThreadLoop,
index_t ThreadGemmADataPerRead_K,
index_t ThreadGemmBDataPerRead_W>
index_t ThreadGemmADataPerRead_E2>
struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
struct MatrixIndex
{
......@@ -29,36 +30,48 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
index_t w;
};
static constexpr index_t KPerThreadLoop = 4;
static constexpr auto E1 = ABlockDesc_E1_K_E2{}.GetLength(I0);
static constexpr auto K = ABlockDesc_E1_K_E2{}.GetLength(I1);
static constexpr auto E2 = ABlockDesc_E1_K_E2{}.GetLength(I2);
static constexpr auto KPerThread = ThreadMatrixC{}.GetLength(I0);
static constexpr auto HPerThread = ThreadMatrixC{}.GetLength(I2);
static constexpr auto WPerThread = ThreadMatrixC{}.GetLength(I3);
static constexpr auto H = BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I2);
static constexpr auto W = BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I3);
static constexpr auto KPerThread = CThreadDesc_K_N_Ho_Wo{}.GetLength(I0);
static constexpr auto HPerThread = CThreadDesc_K_N_Ho_Wo{}.GetLength(I2);
static constexpr auto WPerThread = CThreadDesc_K_N_Ho_Wo{}.GetLength(I3);
static constexpr index_t KPerThreadLoop = KPerThread;
static constexpr auto a_thread_mtx_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<EPerThreadLoop>{}, Number<KPerThreadLoop>{}));
make_tuple(Number<EPerThreadLoop>{}, Number<KPerThreadLoop>{}, Number<E2>{}));
static constexpr auto b_thread_mtx_ = make_naive_tensor_descriptor_packed(make_tuple(
Number<EPerThreadLoop>{}, Number<1>{}, Number<HPerThread>{}, Number<WPerThread>{}));
static constexpr auto b_thread_mtx_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<EPerThreadLoop>{},
Number<1>{},
Number<HPerThread>{},
Number<WPerThread>{},
Number<E2>{}));
static constexpr auto c_thread_mtx_ = make_naive_tensor_descriptor_packed(make_tuple(
Number<KPerThreadLoop>{}, Number<1>{}, Number<HPerThread>{}, Number<WPerThread>{}));
__device__ BlockwiseGemmDlops_km_kn_m0m1n0n1_v3()
: c_thread_begin_mtx_idx_{GetBeginOfThreadMatrixC(get_thread_local_1d_id())},
a_thread_copy_{make_tuple(0, c_thread_begin_mtx_idx_.k * KPerThread)}
: c_thread_begin_mtx_idx_{GetBeginOfCThreadDesc_K_N_Ho_Wo(get_thread_local_1d_id())},
a_thread_copy_{make_tuple(0, c_thread_begin_mtx_idx_.k * KPerThread, 0)}
{
static_assert(BlockMatrixA::IsKnownAtCompileTime() &&
BlockMatrixB::IsKnownAtCompileTime() &&
ThreadMatrixC::IsKnownAtCompileTime(),
static_assert(ABlockDesc_E1_K_E2::IsKnownAtCompileTime() &&
BBlockDesc_E1_N_Ho_Wo_E2::IsKnownAtCompileTime() &&
CThreadDesc_K_N_Ho_Wo::IsKnownAtCompileTime(),
"wrong! Desc should be known at compile-time");
static_assert(BlockMatrixA{}.GetLength(I0) == BlockMatrixB{}.GetLength(I0),
"wrong! K dimension not consistent\n");
static_assert(
ABlockDesc_E1_K_E2{}.GetLength(I0) == BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I0) &&
ABlockDesc_E1_K_E2{}.GetLength(I2) == BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I4),
"wrong! E dimension not consistent\n");
constexpr index_t K = BlockMatrixA{}.GetLength(I1); // A is transposed
constexpr index_t H = BlockMatrixB{}.GetLength(I2);
constexpr index_t W = BlockMatrixB{}.GetLength(I3);
static_assert(E1 % EPerThreadLoop == 0, "");
static_assert(KPerThread % KPerThreadLoop == 0, "");
static_assert(K % KPerThread == 0 && H % HPerThread == 0 && W % WPerThread == 0,
"wrong! Cannot evenly divide work among\n");
......@@ -71,15 +84,15 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
"wrong! wrong blocksize\n");
}
__device__ static constexpr auto GetThreadMatrixCLengths()
__device__ static constexpr auto GetCThreadDesc_K_N_Ho_WoLengths()
{
return Sequence<KPerThread, 1, HPerThread, WPerThread>{};
}
__device__ static MatrixIndex GetBeginOfThreadMatrixC(index_t thread_id)
__device__ static MatrixIndex GetBeginOfCThreadDesc_K_N_Ho_Wo(index_t thread_id)
{
constexpr index_t HPerBlock = BlockMatrixB{}.GetLength(Number<2>{});
constexpr index_t WPerBlock = BlockMatrixB{}.GetLength(Number<3>{});
constexpr index_t HPerBlock = BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I2);
constexpr index_t WPerBlock = BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I3);
constexpr auto num_w_threads = WPerBlock / WPerThread;
constexpr auto num_h_threads = HPerBlock / HPerThread;
......@@ -100,42 +113,37 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
CThreadBuffer& c_thread_buf) const
{
static_assert(
is_same<remove_cvref_t<typename ABlockBuffer::type>, remove_cvref_t<FloatAB>>::value &&
is_same<remove_cvref_t<typename BThreadBuffer::type>, remove_cvref_t<FloatAB>>::value &&
is_same<remove_cvref_t<typename ABlockBuffer::type>, remove_cvref_t<FloatA>>::value &&
is_same<remove_cvref_t<typename BThreadBuffer::type>, remove_cvref_t<FloatB>>::value &&
is_same<remove_cvref_t<typename CThreadBuffer::type>, remove_cvref_t<FloatC>>::value &&
"wrong! inconsistent type");
constexpr auto a_block_mtx = BlockMatrixA{};
constexpr auto EPerBlock = a_block_mtx.GetLength(I0);
static_assert(EPerBlock % EPerThreadLoop == 0, "");
static_assert(KPerThread % KPerThreadLoop == 0, "");
constexpr auto a_block_mtx = ABlockDesc_E1_K_E2{};
// thread A buffer for GEMM
StaticBuffer<AddressSpaceEnum_t::Vgpr, FloatAB, a_thread_mtx_.GetElementSpaceSize(), true>
StaticBuffer<AddressSpaceEnum_t::Vgpr, FloatA, a_thread_mtx_.GetElementSpaceSize(), true>
a_thread_buf;
constexpr auto threadwise_gemm = ThreadwiseGemmDlops_km_kn_mn_v3<FloatAB,
FloatAB,
constexpr auto threadwise_gemm = ThreadwiseGemmDlops_km_kn_mn_v3<FloatA,
FloatB,
FloatC,
decltype(a_thread_mtx_),
decltype(b_thread_mtx_),
decltype(c_thread_mtx_)>{};
static_for<0, EPerBlock, EPerThreadLoop>{}([&](auto e_begin) {
static_for<0, E1, EPerThreadLoop>{}([&](auto e_begin) {
static_for<0, KPerThread, KPerThreadLoop>{}([&](auto k_begin) {
a_thread_copy_.Run(a_block_mtx,
make_tuple(e_begin, k_begin),
make_tuple(e_begin, k_begin, I0),
a_block_buf,
a_thread_mtx_,
make_tuple(I0, I0),
make_tuple(I0, I0, I0),
a_thread_buf);
threadwise_gemm.Run(a_thread_buf,
make_tuple(I0, I0),
make_tuple(I0, I0, I0),
b_thread_buf,
make_tuple(e_begin, I0, I0, I0),
make_tuple(e_begin, I0, I0, I0, I0),
c_thread_buf,
make_tuple(k_begin, I0, I0, I0));
});
......@@ -145,21 +153,22 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
template <typename ABlockSliceMoveStepIdx>
__device__ void MoveABlockSliceWindow(const ABlockSliceMoveStepIdx& a_block_slice_move_step_idx)
{
a_thread_copy_.MoveSrcSliceWindow(BlockMatrixA{}, a_block_slice_move_step_idx);
a_thread_copy_.MoveSrcSliceWindow(ABlockDesc_E1_K_E2{}, a_block_slice_move_step_idx);
}
private:
MatrixIndex c_thread_begin_mtx_idx_;
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
BlockMatrixA,
decltype(a_thread_mtx_),
Sequence<EPerThreadLoop, KPerThreadLoop>,
Sequence<0, 1>,
1,
ThreadGemmADataPerRead_K,
1>;
using AThreadCopy =
ThreadwiseTensorSliceTransfer_v4<FloatA,
FloatB,
ABlockDesc_E1_K_E2,
decltype(a_thread_mtx_),
Sequence<EPerThreadLoop, KPerThreadLoop, E2>,
Sequence<0, 1, 2>,
2,
ThreadGemmADataPerRead_E2,
ThreadGemmADataPerRead_E2>;
AThreadCopy a_thread_copy_;
};
......
composable_kernel/include/tensor_operation/gridwise_gemm_dlops_v2.hpp
View file @ e6a23d8b
......@@ -15,8 +15,8 @@ namespace ck {
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename AGridDesc_E0_E1_K,
typename BGridDesc_E_N_Ho_Wo,
typename AGridDesc_E0_E1_K_E2,
typename BGridDesc_E0_E1_N_Ho_Wo_E2,
typename CGridDesc_K_N_Ho_Wo,
typename CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo,
bool HasMainKBlockLoop,
......@@ -28,8 +28,8 @@ __global__ void
kernel_gemm_dlops_v2(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const AGridDesc_E0_E1_K a_e0_e1_k_grid_desc,
const BGridDesc_E_N_Ho_Wo b_e0_e1_n_ho_wo_grid_desc,
const AGridDesc_E0_E1_K_E2 a_e0_e1_k_e2_grid_desc,
const BGridDesc_E0_E1_N_Ho_Wo_E2 b_e0_e1_n_ho_wo_e2_grid_desc,
const CGridDesc_K_N_Ho_Wo c_k_n_ho_wo_grid_desc,
const CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo
c_blockid_to_k_n_ho_wo_block_cluster_adaptor)
......@@ -43,8 +43,8 @@ __global__ void
p_b_grid,
p_c_grid,
p_shared_block,
a_e0_e1_k_grid_desc,
b_e0_e1_n_ho_wo_grid_desc,
a_e0_e1_k_e2_grid_desc,
b_e0_e1_n_ho_wo_e2_grid_desc,
c_k_n_ho_wo_grid_desc,
integral_constant<bool, HasMainKBlockLoop>{},
integral_constant<bool, HasDoubleTailKBlockLoop>{});
......@@ -56,8 +56,8 @@ __global__ void
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename AGridDesc_E0_E1_K,
typename BGridDesc_E_N_Ho_Wo,
typename AGridDesc_E0_E1_K_E2,
typename BGridDesc_E0_E1_N_Ho_Wo_E2,
typename CGridDesc_K_N_Ho_Wo,
typename CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo,
bool HasMainKBlockLoop,
......@@ -69,18 +69,18 @@ __global__ void
kernel_gemm_dlops_v2(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const void CONSTANT* p_a_e0_e1_k_grid_desc,
const void CONSTANT* p_b_e0_e1_n_ho_wo_grid_desc,
const void CONSTANT* p_a_e0_e1_k_e2_grid_desc,
const void CONSTANT* p_b_e0_e1_n_ho_wo_e2_grid_desc,
const void CONSTANT* p_c_k_n_ho_wo_grid_desc,
const void CONSTANT* p_c_blockid_to_k_n_ho_wo_block_cluster_adaptor)
{
// first cast void CONSTANT void* to void*
// second cast void* to Desc*
// the copy constructor of tensor descriptor doesn't take address_space(4)
const auto a_e0_e1_k_grid_desc = *reinterpret_cast<const AGridDesc_E0_E1_K*>(
cast_pointer_to_generic_address_space(p_a_e0_e1_k_grid_desc));
const auto b_e0_e1_n_ho_wo_grid_desc = *reinterpret_cast<const BGridDesc_E_N_Ho_Wo*>(
cast_pointer_to_generic_address_space(p_b_e0_e1_n_ho_wo_grid_desc));
const auto a_e0_e1_k_e2_grid_desc = *reinterpret_cast<const AGridDesc_E0_E1_K_E2*>(
cast_pointer_to_generic_address_space(p_a_e0_e1_k_e2_grid_desc));
const auto b_e0_e1_n_ho_wo_e2_grid_desc = *reinterpret_cast<const BGridDesc_E0_E1_N_Ho_Wo_E2*>(
cast_pointer_to_generic_address_space(p_b_e0_e1_n_ho_wo_e2_grid_desc));
const auto c_k_n_ho_wo_grid_desc = *reinterpret_cast<const CGridDesc_K_N_Ho_Wo*>(
cast_pointer_to_generic_address_space(p_c_k_n_ho_wo_grid_desc));
......@@ -93,8 +93,8 @@ __global__ void
p_b_grid,
p_c_grid,
p_shared_block,
a_e0_e1_k_grid_desc,
b_e0_e1_n_ho_wo_grid_desc,
a_e0_e1_k_e2_grid_desc,
b_e0_e1_n_ho_wo_e2_grid_desc,
c_k_n_ho_wo_grid_desc,
integral_constant<bool, HasMainKBlockLoop>{},
integral_constant<bool, HasDoubleTailKBlockLoop>{});
......@@ -106,10 +106,11 @@ template <index_t BlockSize,
typename FloatAcc,
typename FloatC,
InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
typename AGlobalDesc_E0_E1_K,
typename BGlobalDesc_E0_E1_N_Ho_Wo,
typename AGlobalDesc_E0_E1_K_E2,
typename BGlobalDesc_E0_E1_N_Ho_Wo_E2,
typename CGlobalDesc_K_N_Ho_Wo,
index_t E1,
index_t E2,
index_t KPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
......@@ -118,13 +119,13 @@ template <index_t BlockSize,
index_t HoPerThread,
index_t WoPerThread,
index_t EPerThread,
typename ABlockTransferThreadSliceLengths_E0_E1_K,
typename ABlockTransferThreadClusterLengths_E0_E1_K,
typename ABlockTransferThreadSliceLengths_E0_E1_K_E2,
typename ABlockTransferThreadClusterLengths_E0_E1_K_E2,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_K,
index_t ABlockTransferDstScalarPerVector_E2,
bool AThreadTransferSrcResetCoordinateAfterRun,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
......@@ -145,20 +146,20 @@ struct GridwiseGemmDlops_km_kn_mn_v3
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
constexpr auto max_lds_align =
math::lcm(Number<ABlockTransferDstScalarPerVector_K>{}, Number<KPerBlock>{});
constexpr auto max_lds_align = Number<ABlockTransferDstScalarPerVector_E2>{};
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_e0_e1_k_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(I1, Number<E1>{}, Number<KPerBlock>{}), max_lds_align);
constexpr auto a_e0_e1_k_e2_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(I1, Number<E1>{}, Number<KPerBlock>{}, Number<E2>{}), max_lds_align);
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size =
math::integer_least_multiple(a_e0_e1_k_block_desc.GetElementSpaceSize(), max_lds_align);
constexpr auto a_block_space_size = math::integer_least_multiple(
a_e0_e1_k_e2_block_desc.GetElementSpaceSize(), max_lds_align);
return a_block_space_size * sizeof(FloatAB);
}
......@@ -168,27 +169,27 @@ struct GridwiseGemmDlops_km_kn_mn_v3
const FloatAB* __restrict__ p_b_global,
FloatC* __restrict__ p_c_global,
FloatAB* __restrict__ p_shared_block,
const AGlobalDesc_E0_E1_K& a_e0_e1_k_global_desc,
const BGlobalDesc_E0_E1_N_Ho_Wo& b_e0_e1_n_ho_wo_global_desc,
const AGlobalDesc_E0_E1_K_E2& a_e0_e1_k_e2_global_desc,
const BGlobalDesc_E0_E1_N_Ho_Wo_E2& b_e0_e1_n_ho_wo_e2_global_desc,
const CGlobalDesc_K_N_Ho_Wo& c_k_n_ho_wo_global_desc,
integral_constant<bool, HasMainKBlockLoop>,
integral_constant<bool, HasDoubleTailKBlockLoop>)
{
const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_a_global, a_e0_e1_k_global_desc.GetElementSpaceSize());
p_a_global, a_e0_e1_k_e2_global_desc.GetElementSpaceSize());
const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_b_global, b_e0_e1_n_ho_wo_global_desc.GetElementSpaceSize());
p_b_global, b_e0_e1_n_ho_wo_e2_global_desc.GetElementSpaceSize());
auto c_global_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_c_global, c_k_n_ho_wo_global_desc.GetElementSpaceSize());
static_assert(E1 % EPerBlock == 0, "");
// const auto E = a_e0_e1_k_global_desc.GetLength(I0);
// const auto K = a_e0_e1_k_global_desc.GetLength(I1);
// const auto E = a_e0_e1_k_e2_global_desc.GetLength(I0);
// const auto K = a_e0_e1_k_e2_global_desc.GetLength(I1);
// const auto N = b_e0_e1_n_ho_wo_global_desc.GetLength(I1);
const auto Ho = b_e0_e1_n_ho_wo_global_desc.GetLength(I3);
const auto Wo = b_e0_e1_n_ho_wo_global_desc.GetLength(I4);
// const auto N = b_e0_e1_n_ho_wo_e2_global_desc.GetLength(I1);
const auto Ho = b_e0_e1_n_ho_wo_e2_global_desc.GetLength(I3);
const auto Wo = b_e0_e1_n_ho_wo_e2_global_desc.GetLength(I4);
// divide block work by [M, N]
#if 1
......@@ -217,39 +218,44 @@ struct GridwiseGemmDlops_km_kn_mn_v3
#endif
// lds max alignment
constexpr auto max_lds_align =
math::lcm(Number<ABlockTransferDstScalarPerVector_K>{}, Number<KPerBlock>{});
constexpr auto max_lds_align = Number<ABlockTransferDstScalarPerVector_E2>{};
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_e0_e1_k_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<I1>{}, Number<E1>{}, Number<KPerBlock>{}), max_lds_align);
constexpr auto a_e1_k_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<EPerBlock>{}, Number<KPerBlock>{}), max_lds_align);
constexpr auto a_e0_e1_k_e2_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<I1>{}, Number<E1>{}, Number<KPerBlock>{}, Number<E2>{}),
max_lds_align);
// B matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto b_e1_n_ho_wo_block_desc = make_naive_tensor_descriptor_packed(make_tuple(
Number<EPerBlock>{}, Number<1>{}, Number<HoPerBlock>{}, Number<WoPerBlock>{}));
constexpr auto b_e1_n_ho_wo_e2_block_desc =
make_naive_tensor_descriptor_packed(make_tuple(Number<EPerBlock>{},
Number<1>{},
Number<HoPerBlock>{},
Number<WoPerBlock>{},
Number<E2>{}));
// c_thread_mtx definition: this is a mess
// TODO:: more elegent way of defining c_thread_mtx
constexpr auto c_k_n_ho_wo_thread_desc = make_naive_tensor_descriptor_packed(make_tuple(
Number<KPerThread>{}, Number<1>{}, Number<HoPerThread>{}, Number<WoPerThread>{}));
constexpr auto a_e1_k_e2_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<EPerBlock>{}, Number<KPerBlock>{}, Number<E2>{}), max_lds_align);
auto blockwise_gemm =
BlockwiseGemmDlops_km_kn_m0m1n0n1_v3<BlockSize,
FloatAB,
FloatAB,
FloatAcc,
decltype(a_e1_k_block_desc),
decltype(b_e1_n_ho_wo_block_desc),
decltype(a_e1_k_e2_block_desc),
decltype(b_e1_n_ho_wo_e2_block_desc),
decltype(c_k_n_ho_wo_thread_desc),
EPerThread,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K>{};
ABlockTransferDstScalarPerVector_E2>{};
auto c_thread_mtx_index = blockwise_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
auto c_thread_mtx_index =
blockwise_gemm.GetBeginOfCThreadDesc_K_N_Ho_Wo(get_thread_local_1d_id());
const auto k_thread_id = c_thread_mtx_index.k;
const auto ho_thread_id = c_thread_mtx_index.h;
......@@ -268,49 +274,53 @@ struct GridwiseGemmDlops_km_kn_mn_v3
auto a_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
InMemoryDataOperationEnum_t::Set,
Sequence<I1, E1, KPerBlock>,
ABlockTransferThreadSliceLengths_E0_E1_K,
ABlockTransferThreadClusterLengths_E0_E1_K,
Sequence<I1, E1, KPerBlock, E2>,
ABlockTransferThreadSliceLengths_E0_E1_K_E2,
ABlockTransferThreadClusterLengths_E0_E1_K_E2,
ABlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(a_e0_e1_k_global_desc),
decltype(a_e0_e1_k_block_desc),
decltype(a_e0_e1_k_e2_global_desc),
decltype(a_e0_e1_k_e2_block_desc),
ABlockTransferSrcAccessOrder,
Sequence<0, 1, 2>, // ABlockTransferDstAccessOrder
Sequence<0, 1, 2, 3>, // ABlockTransferDstAccessOrder
ABlockTransferSrcVectorDim,
2, // ABlockTransferDstVectorDim
3, // ABlockTransferDstVectorDim
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K,
ABlockTransferDstScalarPerVector_E2,
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true>(a_e0_e1_k_global_desc,
make_multi_index(0, 0, k_block_data_on_global),
a_e0_e1_k_block_desc,
make_multi_index(0, 0, 0));
constexpr auto a_block_slice_copy_step = make_multi_index(I1, 0, 0);
constexpr auto b_e0_e1_n_ho_wo_thread_desc = make_naive_tensor_descriptor_packed(make_tuple(
I1, Number<EPerBlock>{}, Number<1>{}, Number<HoPerThread>{}, Number<WoPerThread>{}));
auto b_threadwise_transfer =
ThreadwiseTensorSliceTransfer_v2<FloatAB,
FloatAB,
decltype(b_e0_e1_n_ho_wo_global_desc),
decltype(b_e0_e1_n_ho_wo_thread_desc),
Sequence<I1, EPerBlock, 1, HoPerThread, WoPerThread>,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
1,
true>(
b_e0_e1_n_ho_wo_global_desc,
make_multi_index(0, 0, 0, ho_thread_data_on_global, wo_thread_data_on_global));
true>(a_e0_e1_k_e2_global_desc,
make_multi_index(0, 0, k_block_data_on_global, 0),
a_e0_e1_k_e2_block_desc,
make_multi_index(0, 0, 0, 0));
constexpr auto a_block_slice_copy_step = make_multi_index(I1, 0, 0, 0);
constexpr auto b_e0_e1_n_ho_wo_e2_thread_desc =
make_naive_tensor_descriptor_packed(make_tuple(I1,
Number<EPerBlock>{},
Number<1>{},
Number<HoPerThread>{},
Number<WoPerThread>{},
Number<E2>{}));
auto b_threadwise_transfer = ThreadwiseTensorSliceTransfer_v2<
FloatAB,
FloatAB,
decltype(b_e0_e1_n_ho_wo_e2_global_desc),
decltype(b_e0_e1_n_ho_wo_e2_thread_desc),
Sequence<I1, EPerBlock, 1, HoPerThread, WoPerThread, E2>,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
1,
true>(b_e0_e1_n_ho_wo_e2_global_desc,
make_multi_index(0, 0, 0, ho_thread_data_on_global, wo_thread_data_on_global, 0));
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
p_shared_block, a_e0_e1_k_block_desc.GetElementSpaceSize());
p_shared_block, a_e0_e1_k_e2_block_desc.GetElementSpaceSize());
// register allocation for output
StaticBuffer<AddressSpaceEnum_t::Vgpr,
......@@ -325,20 +335,20 @@ struct GridwiseGemmDlops_km_kn_mn_v3
Sequence<KPerThread, 1, HoPerThread, WoPerThread>>{}
.Run(c_k_n_ho_wo_thread_desc, make_tuple(I0, I0, I0, I0), c_thread_buf, FloatAcc{0});
constexpr auto b_thread_slice_copy_step = make_multi_index(0, EPerBlock, 0, 0, 0);
constexpr auto b_thread_slice_copy_step = make_multi_index(0, EPerBlock, 0, 0, 0, 0);
// hack to control index calculation when iterating over A and B matrix for threadwise copy
constexpr auto a_e0_e1_k_global_step_hacks = AGlobalStepHacks{};
constexpr auto b_e0_e1_n_ho_wo_global_step_hacks = BGlobalStepHacks{};
constexpr auto a_e0_e1_k_e2_global_step_hacks = AGlobalStepHacks{};
constexpr auto b_e0_e1_n_ho_wo_e2_global_step_hacks = BGlobalStepHacks{};
// double regsiter buffer for b
StaticBuffer<AddressSpaceEnum_t::Vgpr,
FloatAB,
b_e0_e1_n_ho_wo_thread_desc.GetElementSpaceSize(),
b_e0_e1_n_ho_wo_e2_thread_desc.GetElementSpaceSize(),
true>
b_thread_even_buf, b_thread_odd_buf;
const auto E0 = b_e0_e1_n_ho_wo_global_desc.GetLength(I0);
const auto E0 = b_e0_e1_n_ho_wo_e2_global_desc.GetLength(I0);
index_t e0_block_data_begin = 0;
......@@ -347,16 +357,16 @@ struct GridwiseGemmDlops_km_kn_mn_v3
// LDS double buffer: preload data
{
a_blockwise_copy.RunRead(
a_e0_e1_k_global_desc, a_global_buf, a_e0_e1_k_global_step_hacks);
a_e0_e1_k_e2_global_desc, a_global_buf, a_e0_e1_k_e2_global_step_hacks);
b_threadwise_transfer.Run(b_e0_e1_n_ho_wo_global_desc,
b_threadwise_transfer.Run(b_e0_e1_n_ho_wo_e2_global_desc,
b_global_buf,
b_e0_e1_n_ho_wo_thread_desc,
make_tuple(I0, I0, I0, I0, I0),
b_e0_e1_n_ho_wo_e2_thread_desc,
make_tuple(I0, I0, I0, I0, I0, I0),
b_thread_even_buf,
b_e0_e1_n_ho_wo_global_step_hacks);
b_e0_e1_n_ho_wo_e2_global_step_hacks);
a_blockwise_copy.RunWrite(a_e0_e1_k_block_desc, a_block_buf);
a_blockwise_copy.RunWrite(a_e0_e1_k_e2_block_desc, a_block_buf);
}
__syncthreads();
......@@ -370,36 +380,36 @@ struct GridwiseGemmDlops_km_kn_mn_v3
do
{
// even iteration
b_threadwise_transfer.MoveSrcSliceWindow(b_e0_e1_n_ho_wo_global_desc,
b_threadwise_transfer.MoveSrcSliceWindow(b_e0_e1_n_ho_wo_e2_global_desc,
b_thread_slice_copy_step);
b_threadwise_transfer.Run(b_e0_e1_n_ho_wo_global_desc,
b_threadwise_transfer.Run(b_e0_e1_n_ho_wo_e2_global_desc,
b_global_buf,
b_e0_e1_n_ho_wo_thread_desc,
make_tuple(I0, I0, I0, I0, I0),
b_e0_e1_n_ho_wo_e2_thread_desc,
make_tuple(I0, I0, I0, I0, I0, I0),
b_thread_odd_buf,
b_e0_e1_n_ho_wo_global_step_hacks);
b_e0_e1_n_ho_wo_e2_global_step_hacks);
// LDS double buffer: GEMM on current data
// TODO: @Zhang Jing: blockwise gemm should be able to move slice window
blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
blockwise_gemm.MoveABlockSliceWindow(make_tuple(EPerBlock, 0));
blockwise_gemm.MoveABlockSliceWindow(make_tuple(EPerBlock, 0, 0));
b_threadwise_transfer.MoveSrcSliceWindow(b_e0_e1_n_ho_wo_global_desc,
b_threadwise_transfer.MoveSrcSliceWindow(b_e0_e1_n_ho_wo_e2_global_desc,
b_thread_slice_copy_step);
b_threadwise_transfer.Run(b_e0_e1_n_ho_wo_global_desc,
b_threadwise_transfer.Run(b_e0_e1_n_ho_wo_e2_global_desc,
b_global_buf,
b_e0_e1_n_ho_wo_thread_desc,
make_tuple(I0, I0, I0, I0, I0),
b_e0_e1_n_ho_wo_e2_thread_desc,
make_tuple(I0, I0, I0, I0, I0, I0),
b_thread_even_buf,
b_e0_e1_n_ho_wo_global_step_hacks);
b_e0_e1_n_ho_wo_e2_global_step_hacks);
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(a_block_buf, b_thread_odd_buf, c_thread_buf);
blockwise_gemm.MoveABlockSliceWindow(make_tuple(EPerBlock, 0));
blockwise_gemm.MoveABlockSliceWindow(make_tuple(EPerBlock, 0, 0));
e1_block_data_begin += 2 * EPerBlock;
......@@ -409,20 +419,20 @@ struct GridwiseGemmDlops_km_kn_mn_v3
// LDS double buffer: tail
if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
{
b_threadwise_transfer.MoveSrcSliceWindow(b_e0_e1_n_ho_wo_global_desc,
b_threadwise_transfer.MoveSrcSliceWindow(b_e0_e1_n_ho_wo_e2_global_desc,
b_thread_slice_copy_step);
b_threadwise_transfer.Run(b_e0_e1_n_ho_wo_global_desc,
b_threadwise_transfer.Run(b_e0_e1_n_ho_wo_e2_global_desc,
b_global_buf,
b_e0_e1_n_ho_wo_thread_desc,
make_tuple(I0, I0, I0, I0, I0),
b_e0_e1_n_ho_wo_e2_thread_desc,
make_tuple(I0, I0, I0, I0, I0, I0),
b_thread_odd_buf,
b_e0_e1_n_ho_wo_global_step_hacks);
b_e0_e1_n_ho_wo_e2_global_step_hacks);
// LDS double buffer: GEMM on 2nd-last data
blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
blockwise_gemm.MoveABlockSliceWindow(make_tuple(EPerBlock, 0));
blockwise_gemm.MoveABlockSliceWindow(make_tuple(EPerBlock, 0, 0));
// LDS double buffer: GEMM on last data
blockwise_gemm.Run(a_block_buf, b_thread_odd_buf, c_thread_buf);
......@@ -433,12 +443,13 @@ struct GridwiseGemmDlops_km_kn_mn_v3
blockwise_gemm.Run(a_block_buf, b_thread_even_buf, c_thread_buf);
}
a_blockwise_copy.MoveSrcSliceWindow(
a_e0_e1_k_global_desc, a_block_slice_copy_step, AGlobalMoveSliceWindowStepHacks{});
a_blockwise_copy.MoveSrcSliceWindow(a_e0_e1_k_e2_global_desc,
a_block_slice_copy_step,
AGlobalMoveSliceWindowStepHacks{});
blockwise_gemm.MoveABlockSliceWindow(make_tuple(-(E1 - EPerBlock), 0));
blockwise_gemm.MoveABlockSliceWindow(make_tuple(-(E1 - EPerBlock), 0, 0));
b_threadwise_transfer.MoveSrcSliceWindow(b_e0_e1_n_ho_wo_global_desc,
b_threadwise_transfer.MoveSrcSliceWindow(b_e0_e1_n_ho_wo_e2_global_desc,
b_thread_slice_copy_step);
e0_block_data_begin += 1;
......
composable_kernel/include/tensor_operation/threadwise_gemm_dlops_v3.hpp
View file @ e6a23d8b
......@@ -9,16 +9,17 @@ namespace ck {
// C[M, N] += transpose(A[K, M]) * B[K, N]
// Element of matrix can be vectorized data
// Assume:
// 1. AThreadDesc_E_K, BThreadDesc_E_N_Ho_Wo, CThreadDesc_K_N_Ho_Wo are known at compile-time
// 1. AThreadDesc_E1_K_E2, BThreadDesc_E1_N_Ho_Wo_E2, CThreadDesc_K_N_Ho_Wo are known at
// compile-time
// 2. AOriginIdx, BOriginIdx, COriginIdx are known at compile-time
template <typename FloatA,
typename FloatB,
typename FloatC,
typename AThreadDesc_E_K,
typename BThreadDesc_E_N_Ho_Wo,
typename AThreadDesc_E1_K_E2,
typename BThreadDesc_E1_N_Ho_Wo_E2,
typename CThreadDesc_K_N_Ho_Wo,
typename enable_if<AThreadDesc_E_K::IsKnownAtCompileTime() &&
BThreadDesc_E_N_Ho_Wo::IsKnownAtCompileTime() &&
typename enable_if<AThreadDesc_E1_K_E2::IsKnownAtCompileTime() &&
BThreadDesc_E1_N_Ho_Wo_E2::IsKnownAtCompileTime() &&
CThreadDesc_K_N_Ho_Wo::IsKnownAtCompileTime(),
bool>::type = false>
struct ThreadwiseGemmDlops_km_kn_mn_v3
......@@ -38,8 +39,8 @@ struct ThreadwiseGemmDlops_km_kn_mn_v3
COriginIdx)
{
static_assert(AThreadDesc_E_K::IsKnownAtCompileTime() &&
BThreadDesc_E_N_Ho_Wo::IsKnownAtCompileTime() &&
static_assert(AThreadDesc_E1_K_E2::IsKnownAtCompileTime() &&
BThreadDesc_E1_N_Ho_Wo_E2::IsKnownAtCompileTime() &&
CThreadDesc_K_N_Ho_Wo::IsKnownAtCompileTime(),
"wrong! Desc should be known at compile-time");
......@@ -54,18 +55,19 @@ struct ThreadwiseGemmDlops_km_kn_mn_v3
is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
"wrong! inconsistent type");
constexpr index_t Vec = 2;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto E = AThreadDesc_E_K{}.GetLength(I0);
constexpr auto K = AThreadDesc_E_K{}.GetLength(I1);
constexpr auto E1 = AThreadDesc_E1_K_E2{}.GetLength(I0);
constexpr auto K = AThreadDesc_E1_K_E2{}.GetLength(I1);
constexpr auto E2 = AThreadDesc_E1_K_E2{}.GetLength(I2);
static_assert(E1 == 4 && E2 == 4, "");
constexpr auto H = BThreadDesc_E_N_Ho_Wo{}.GetLength(I2);
constexpr auto W = BThreadDesc_E_N_Ho_Wo{}.GetLength(I3);
constexpr auto H = BThreadDesc_E1_N_Ho_Wo_E2{}.GetLength(I2);
constexpr auto W = BThreadDesc_E1_N_Ho_Wo_E2{}.GetLength(I3);
constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
......@@ -74,22 +76,23 @@ struct ThreadwiseGemmDlops_km_kn_mn_v3
static_for<0, K, 1>{}([&](auto k) {
static_for<0, H, 1>{}([&](auto h) {
static_for<0, W, 1>{}([&](auto w) {
static_for<0, E, Vec>{}([&](auto e) {
vector_type<FloatA, Vec> a_vec;
vector_type<FloatB, Vec> b_vec;
static_for<0, E1, 1>{}([&](auto e) {
vector_type<FloatA, E2> a_vec;
vector_type<FloatB, E2> b_vec;
static_for<0, Vec, 1>{}([&](auto v) {
constexpr index_t a_offset = AThreadDesc_E_K{}.CalculateOffset(
a_origin_idx + make_tuple(e + v, k));
constexpr index_t b_offset = BThreadDesc_E_N_Ho_Wo{}.CalculateOffset(
b_origin_idx + make_tuple(e + v, 0, h, w));
static_for<0, E2, 1>{}([&](auto v) {
constexpr index_t a_offset = AThreadDesc_E1_K_E2{}.CalculateOffset(
a_origin_idx + make_tuple(e, k, v));
constexpr index_t b_offset =
BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
b_origin_idx + make_tuple(e, 0, h, w, v));
a_vec.template AsType<FloatA>()(v) = a_buf[Number<a_offset>{}];
b_vec.template AsType<FloatB>()(v) = b_buf[Number<b_offset>{}];
});
using a_vector_t = typename vector_type<FloatA, Vec>::type;
using b_vector_t = typename vector_type<FloatB, Vec>::type;
using a_vector_t = typename vector_type<FloatA, E2>::type;
using b_vector_t = typename vector_type<FloatB, E2>::type;
constexpr index_t c_offset = CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(
c_origin_idx + make_tuple(k, 0, h, w));
......
host/driver_offline/include/device_convolution_forward_implicit_gemm_v5r1_dlops_nchw_kcyx_nkhw.hpp
View file @ e6a23d8b
......@@ -102,26 +102,27 @@ void device_convolution_forward_implicit_gemm_v5r1_dlops_nchw_kcyx_nkhw(
constexpr index_t BlockSize = 64;
constexpr index_t KPerBlock = 16;
constexpr index_t HoPerBlock = 32;
constexpr index_t HoPerBlock = 8;
constexpr index_t WoPerBlock = 8;
constexpr index_t E1 = 16;
constexpr index_t EPerBlock = 16;
constexpr index_t E1 = 4;
constexpr index_t E2 = 4;
constexpr index_t EPerBlock = 4;
constexpr index_t KPerThread = KPerBlock;
constexpr index_t KPerThread = 4;
constexpr index_t HoPerThread = 2;
constexpr index_t WoPerThread = 2;
constexpr index_t EPerThread = EPerBlock;
constexpr index_t EPerThread = 4;
using ABlockTransferThreadSliceLengths_E0_E1_K = Sequence<1, 4, 1>;
using ABlockTransferThreadClusterLengths_E0_E1_K = Sequence<1, 4, 16>;
using ABlockTransferThreadSliceLengths_E0_E1_K_E2 = Sequence<1, 1, 1, 4>;
using ABlockTransferThreadClusterLengths_E0_E1_K_E2 = Sequence<1, 4, 16, 1>;
constexpr index_t ABlockTransferSrcScalarPerVector_E = 4;
constexpr index_t ABlockTransferDstScalarPerVector_K = 1;
constexpr index_t ABlockTransferSrcScalarPerVector_E2 = 1;
constexpr index_t ABlockTransferDstScalarPerVector_E2 = 1;
constexpr index_t BThreadTransferSrcScalarPerVector_E = 4;
constexpr index_t BThreadTransferSrcScalarPerVector_E2 = 1;
constexpr index_t CThreadTransferDstScalarPerVector_K = 4;
constexpr index_t CThreadTransferDstScalarPerVector_K = 1;
#endif
constexpr auto conv_driver =
......@@ -131,6 +132,7 @@ void device_convolution_forward_implicit_gemm_v5r1_dlops_nchw_kcyx_nkhw(
TAcc,
TOut,
E1,
E2,
KPerBlock,
HoPerBlock,
WoPerBlock,
......@@ -139,11 +141,11 @@ void device_convolution_forward_implicit_gemm_v5r1_dlops_nchw_kcyx_nkhw(
HoPerThread,
WoPerThread,
EPerThread,
ABlockTransferThreadSliceLengths_E0_E1_K,
ABlockTransferThreadClusterLengths_E0_E1_K,
ABlockTransferSrcScalarPerVector_E,
ABlockTransferDstScalarPerVector_K,
BThreadTransferSrcScalarPerVector_E,
ABlockTransferThreadSliceLengths_E0_E1_K_E2,
ABlockTransferThreadClusterLengths_E0_E1_K_E2,
ABlockTransferSrcScalarPerVector_E2,
ABlockTransferDstScalarPerVector_E2,
BThreadTransferSrcScalarPerVector_E2,
CThreadTransferDstScalarPerVector_K>{};
const auto ave_time =
......
host/driver_offline/include/driver_convolution_forward_implicit_gemm_v5r1_dlops_nchw_kcyx_nkhw_outpad.hpp
View file @ e6a23d8b
......@@ -11,6 +11,7 @@ template <ck::index_t BlockSize,
typename FloatAcc,
typename FloatC,
ck::index_t E1,
ck::index_t E2,
ck::index_t KPerBlock,
ck::index_t HoPerBlock,
ck::index_t WoPerBlock,
......@@ -19,11 +20,11 @@ template <ck::index_t BlockSize,
ck::index_t HoPerThread,
ck::index_t WoPerThread,
ck::index_t EPerThread,
typename ABlockTransferThreadSliceLengths_E_K,
typename ABlockTransferThreadClusterLengths_E_K,
ck::index_t ABlockTransferSrcScalarPerVector_E,
ck::index_t ABlockTransferDstScalarPerVector_K,
ck::index_t BThreadTransferSrcScalarPerVector_E,
typename ABlockTransferThreadSliceLengths_E0_E1_K_E2,
typename ABlockTransferThreadClusterLengths_E0_E1_K_E2,
ck::index_t ABlockTransferSrcScalarPerVector_E2,
ck::index_t ABlockTransferDstScalarPerVector_E2,
ck::index_t BThreadTransferSrcScalarPerVector_E2,
ck::index_t CThreadTransferDstScalarPerVector_K>
struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outpad
{
......@@ -93,7 +94,7 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
<< std::endl;
const auto E = C0 * Y * X * C1;
const auto E0 = E / E1;
const auto E0 = E / (E1 * E2);
// weight tensor
const auto a_e_k_grid_desc = transform_tensor_descriptor(
......@@ -103,11 +104,12 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
const auto a_e0_e1_k_grid_desc = transform_tensor_descriptor(
a_e_k_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(E0, E1)), make_pass_through_transform(K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2>{}));
const auto a_e0_e1_k_e2_grid_desc =
transform_tensor_descriptor(a_e_k_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(E0, E1, E2)),
make_pass_through_transform(K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// input tensor
const auto in_n_c0_hip_wip_c1_global_desc = transform_tensor_descriptor(
......@@ -141,14 +143,14 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
make_tuple(Sequence<1, 2, 4, 6>{}, Sequence<0>{}, Sequence<3>{}, Sequence<5>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto b_e0_e1_n_ho_wo_grid_desc = transform_tensor_descriptor(
const auto b_e0_e1_n_ho_wo_e2_grid_desc = transform_tensor_descriptor(
b_e_n_ho_wo_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(E0, E1)),
make_tuple(make_unmerge_transform(make_tuple(E0, E1, E2)),
make_pass_through_transform(N),
make_pass_through_transform(Hop),
make_pass_through_transform(Wop)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
make_tuple(Sequence<0, 1, 5>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
// output tensor
const auto c_k_n_hop_wop_grid_desc = transform_tensor_descriptor(
......@@ -169,27 +171,33 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
}
// hack to control index calculation when iterating over a_k_m_global tensor
constexpr auto a_e0_e1_k_global_step_hacks = make_tuple(
make_tuple(
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}),
make_tuple(
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}));
constexpr auto a_e0_e1_k_e2_global_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}));
constexpr auto a_e0_e1_k_global_move_slice_window_step_hack = Sequence<0, 0, 0, 0, 0>{};
constexpr auto a_e0_e1_k_e2_global_move_slice_window_step_hack = Sequence<0, 0, 0, 0, 0>{};
constexpr auto b_e0_e1_n_ho_wo_global_step_hacks = make_tuple(
constexpr auto b_e0_e1_n_ho_wo_e2_global_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}));
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0>{}));
constexpr auto b_e0_e1_n_ho_wo_global_move_slice_window_step_hack =
constexpr auto b_e0_e1_n_ho_wo_e2_global_move_slice_window_step_hack =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0>{};
// hack to control index calculation when iterating over c_m0_m1_n0_n1_global tensor
......@@ -211,10 +219,11 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
FloatAcc,
FloatC,
InMemoryDataOperationEnum_t::Set,
decltype(a_e0_e1_k_grid_desc),
decltype(b_e0_e1_n_ho_wo_grid_desc),
decltype(a_e0_e1_k_e2_grid_desc),
decltype(b_e0_e1_n_ho_wo_e2_grid_desc),
decltype(c_k_n_hop_wop_grid_desc),
E1,
E2,
KPerBlock,
HoPerBlock,
WoPerBlock,
......@@ -223,31 +232,31 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
HoPerThread,
WoPerThread,
EPerThread,
ABlockTransferThreadSliceLengths_E_K,
ABlockTransferThreadClusterLengths_E_K,
Sequence<2, 0, 1>,
Sequence<2, 0, 1>,
1,
ABlockTransferSrcScalarPerVector_E,
ABlockTransferDstScalarPerVector_K,
ABlockTransferThreadSliceLengths_E0_E1_K_E2,
ABlockTransferThreadClusterLengths_E0_E1_K_E2,
Sequence<2, 0, 1, 3>,
Sequence<2, 0, 1, 3>,
3,
ABlockTransferSrcScalarPerVector_E2,
ABlockTransferDstScalarPerVector_E2,
false, // don't move back src coordinate after threadwise copy
Sequence<0, 2, 3, 4, 1>,
1,
BThreadTransferSrcScalarPerVector_E,
Sequence<0, 2, 3, 4, 1, 5>,
5,
BThreadTransferSrcScalarPerVector_E2,
false, // don't move back src coordinate after threadwise copy, which will be fused with
// MoveSrcSliceWindow() to save addr computation
Sequence<0, 2, 3, 1>,
Sequence<2, 3, 1, 0>,
0,
CThreadTransferDstScalarPerVector_K,
decltype(a_e0_e1_k_global_step_hacks),
decltype(b_e0_e1_n_ho_wo_global_step_hacks),
decltype(a_e0_e1_k_e2_global_step_hacks),
decltype(b_e0_e1_n_ho_wo_e2_global_step_hacks),
decltype(c_k_n_ho_wo_global_tensor_step_hacks),
decltype(a_e0_e1_k_global_move_slice_window_step_hack),
decltype(b_e0_e1_n_ho_wo_global_move_slice_window_step_hack)>;
decltype(a_e0_e1_k_e2_global_move_slice_window_step_hack),
decltype(b_e0_e1_n_ho_wo_e2_global_move_slice_window_step_hack)>;
using AGridDesc_E0_E1_K = decltype(a_e0_e1_k_grid_desc);
using BGridDesc_E0_E1_N_Ho_Wo = decltype(b_e0_e1_n_ho_wo_grid_desc);
using CGridDesc_K_N_Ho_Wo = decltype(c_k_n_hop_wop_grid_desc);
using AGridDesc_E0_E1_K_E2 = decltype(a_e0_e1_k_e2_grid_desc);
using BGridDesc_E0_E1_N_Ho_Wo_E2 = decltype(b_e0_e1_n_ho_wo_e2_grid_desc);
using CGridDesc_K_N_Ho_Wo = decltype(c_k_n_hop_wop_grid_desc);
const auto grid_size = (K / KPerBlock) * (Hop / HoPerBlock) * (Wop / WoPerBlock) * N;
......@@ -276,8 +285,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
kernel_gemm_dlops_v2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo>,
remove_reference_t<AGridDesc_E0_E1_K_E2>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo_E2>,
remove_reference_t<CGridDesc_K_N_Ho_Wo>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo>,
true,
......@@ -291,8 +300,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
p_a_grid,
p_b_grid,
p_c_grid,
a_e0_e1_k_grid_desc,
b_e0_e1_n_ho_wo_grid_desc,
a_e0_e1_k_e2_grid_desc,
b_e0_e1_n_ho_wo_e2_grid_desc,
c_k_n_hop_wop_grid_desc,
c_blockid_to_k_n_ho_wo_block_cluster_adaptor);
}
......@@ -302,8 +311,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
kernel_gemm_dlops_v2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo>,
remove_reference_t<AGridDesc_E0_E1_K_E2>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo_E2>,
remove_reference_t<CGridDesc_K_N_Ho_Wo>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo>,
true,
......@@ -317,8 +326,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
p_a_grid,
p_b_grid,
p_c_grid,
a_e0_e1_k_grid_desc,
b_e0_e1_n_ho_wo_grid_desc,
a_e0_e1_k_e2_grid_desc,
b_e0_e1_n_ho_wo_e2_grid_desc,
c_k_n_hop_wop_grid_desc,
c_blockid_to_k_n_ho_wo_block_cluster_adaptor);
}
......@@ -328,8 +337,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
kernel_gemm_dlops_v2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo>,
remove_reference_t<AGridDesc_E0_E1_K_E2>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo_E2>,
remove_reference_t<CGridDesc_K_N_Ho_Wo>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo>,
false,
......@@ -343,8 +352,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
p_a_grid,
p_b_grid,
p_c_grid,
a_e0_e1_k_grid_desc,
b_e0_e1_n_ho_wo_grid_desc,
a_e0_e1_k_e2_grid_desc,
b_e0_e1_n_ho_wo_e2_grid_desc,
c_k_n_hop_wop_grid_desc,
c_blockid_to_k_n_ho_wo_block_cluster_adaptor);
}
......@@ -354,8 +363,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
kernel_gemm_dlops_v2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo>,
remove_reference_t<AGridDesc_E0_E1_K_E2>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo_E2>,
remove_reference_t<CGridDesc_K_N_Ho_Wo>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo>,
false,
......@@ -369,22 +378,22 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
p_a_grid,
p_b_grid,
p_c_grid,
a_e0_e1_k_grid_desc,
b_e0_e1_n_ho_wo_grid_desc,
a_e0_e1_k_e2_grid_desc,
b_e0_e1_n_ho_wo_e2_grid_desc,
c_k_n_hop_wop_grid_desc,
c_blockid_to_k_n_ho_wo_block_cluster_adaptor);
}
return ave_time;
#elif CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER
DeviceMem a_e0_e1_k_grid_desc_dev_buf(sizeof(AGridDesc_E0_E1_K));
DeviceMem b_e0_e1_n_ho_wo_grid_desc_dev_buf(sizeof(BGridDesc_E0_E1_N_Ho_Wo));
DeviceMem a_e0_e1_k_e2_grid_desc_dev_buf(sizeof(AGridDesc_E0_E1_K_E2));
DeviceMem b_e0_e1_n_ho_wo_e2_grid_desc_dev_buf(sizeof(BGridDesc_E0_E1_N_Ho_Wo_E2));
DeviceMem c_k_n_hop_wop_grid_desc_dev_buf(sizeof(CGridDesc_K_N_Ho_Wo));
DeviceMem c_blockid_to_k_n_ho_wo_block_cluster_adaptor_dev_buf(
sizeof(CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo));
a_e0_e1_k_grid_desc_dev_buf.ToDevice(&a_e0_e1_k_grid_desc);
b_e0_e1_n_ho_wo_grid_desc_dev_buf.ToDevice(&b_e0_e1_n_ho_wo_grid_desc);
a_e0_e1_k_e2_grid_desc_dev_buf.ToDevice(&a_e0_e1_k_e2_grid_desc);
b_e0_e1_n_ho_wo_e2_grid_desc_dev_buf.ToDevice(&b_e0_e1_n_ho_wo_e2_grid_desc);
c_k_n_hop_wop_grid_desc_dev_buf.ToDevice(&c_k_n_hop_wop_grid_desc);
c_blockid_to_k_n_ho_wo_block_cluster_adaptor_dev_buf.ToDevice(
&c_blockid_to_k_n_ho_wo_block_cluster_adaptor);
......@@ -397,8 +406,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
kernel_gemm_dlops_v2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo>,
remove_reference_t<AGridDesc_E0_E1_K_E2>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo_E2>,
remove_reference_t<CGridDesc_K_N_Ho_Wo>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo>,
true,
......@@ -414,9 +423,9 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
p_b_grid,
p_c_grid,
cast_pointer_to_constant_address_space(
a_e0_e1_k_grid_desc_dev_buf.GetDeviceBuffer()),
a_e0_e1_k_e2_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
b_e0_e1_n_ho_wo_grid_desc_dev_buf.GetDeviceBuffer()),
b_e0_e1_n_ho_wo_e2_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
c_k_n_hop_wop_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
......@@ -428,8 +437,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
kernel_gemm_dlops_v2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo>,
remove_reference_t<AGridDesc_E0_E1_K_E2>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo_E2>,
remove_reference_t<CGridDesc_K_N_Ho_Wo>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo>,
true,
......@@ -445,9 +454,9 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
p_b_grid,
p_c_grid,
cast_pointer_to_constant_address_space(
a_e0_e1_k_grid_desc_dev_buf.GetDeviceBuffer()),
a_e0_e1_k_e2_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
b_e0_e1_n_ho_wo_grid_desc_dev_buf.GetDeviceBuffer()),
b_e0_e1_n_ho_wo_e2_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
c_k_n_hop_wop_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
......@@ -459,8 +468,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
kernel_gemm_dlops_v2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo>,
remove_reference_t<AGridDesc_E0_E1_K_E2>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo_E2>,
remove_reference_t<CGridDesc_K_N_Ho_Wo>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo>,
false,
......@@ -476,9 +485,9 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
p_b_grid,
p_c_grid,
cast_pointer_to_constant_address_space(
a_e0_e1_k_grid_desc_dev_buf.GetDeviceBuffer()),
a_e0_e1_k_e2_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
b_e0_e1_n_ho_wo_grid_desc_dev_buf.GetDeviceBuffer()),
b_e0_e1_n_ho_wo_e2_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
c_k_n_hop_wop_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
......@@ -490,8 +499,8 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
kernel_gemm_dlops_v2<GridwiseGemm,
FloatAB,
FloatC,
remove_reference_t<AGridDesc_E0_E1_K>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo>,
remove_reference_t<AGridDesc_E0_E1_K_E2>,
remove_reference_t<BGridDesc_E0_E1_N_Ho_Wo_E2>,
remove_reference_t<CGridDesc_K_N_Ho_Wo>,
remove_reference_t<CBlockIdToBlockClusterAdaptor_K_N_Ho_Wo>,
false,
......@@ -507,9 +516,9 @@ struct DriverDynamicConvolutionForwardImplicitGemmDlops_v5r1_nchw_kcyx_nkhw_outp
p_b_grid,
p_c_grid,
cast_pointer_to_constant_address_space(
a_e0_e1_k_grid_desc_dev_buf.GetDeviceBuffer()),
a_e0_e1_k_e2_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
b_e0_e1_n_ho_wo_grid_desc_dev_buf.GetDeviceBuffer()),
b_e0_e1_n_ho_wo_e2_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
c_k_n_hop_wop_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
......
script/run.sh
View file @ e6a23d8b
......@@ -52,6 +52,7 @@ REPEAT=$6
#./host/driver_online/conv_fwd_driver_online $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 128 128 192 3 3 71 71 2 2 1 1 1 1 1 1
./host/driver_offline/conv_fwd_driver_offline $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 1 16 16 3 3 1080 1920 1 1 1 1 1 1 1 1
#./host/driver_offline/conv_fwd_driver_offline $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 1 16 16 1 1 8 8 1 1 1 1 1 1 1 1
################################################ layout algo verify init log repeat M___ N___ K___
#./host/driver_offline/gemm_driver_offline $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 960 1024 1024
......
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