Skip to content

GitLab

Unverified Commit f8a6c69c authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge branch 'develop' into mi300

parents 56599d67 52f64967
Hide whitespace changes
Inline Side-by-side
Showing with 6683 additions and 488 deletions
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
View file @ f8a6c69c
......@@ -13,7 +13,7 @@
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp"
#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
......@@ -25,15 +25,17 @@ namespace device {
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename D0sPointer,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename AccElementwiseOperation,
typename C0DEElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation,
typename C1DEElementwiseOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename B1GridDesc_BK0_N_BK1,
typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
typename Block2CTileMap,
typename ComputeBasePtrOfStridedBatch,
typename C0MatrixMask,
......@@ -47,16 +49,19 @@ __global__ void
const FloatAB* __restrict__ p_b_grid,
const FloatAB* __restrict__ p_b1_grid,
FloatC* __restrict__ p_c_grid,
D0sPointer p_d0s_grid,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const AccElementwiseOperation acc_element_op,
const C0DEElementwiseOperation c0de_element_op,
const B1ElementwiseOperation b1_element_op,
const CElementwiseOperation c_element_op,
const C1DEElementwiseOperation c1de_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 B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c1_grid_desc_mblock_mperblock_nblock_nperblock,
const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
const Block2CTileMap block_2_ctile_map,
const index_t batch_count,
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
......@@ -78,20 +83,28 @@ __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)));
static_for<0, p_d0s_grid.Size(), 1>{}([&](auto In) {
const long_index_t d0_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetD0BasePtr(g_idx, In)));
p_d0s_grid(In) = p_d0s_grid(In) + d0_batch_offset;
});
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset,
p_b1_grid + b1_batch_offset,
p_c_grid + c_batch_offset,
p_d0s_grid,
p_shared,
a_element_op,
b_element_op,
acc_element_op,
c0de_element_op,
b1_element_op,
c_element_op,
c1de_element_op,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
b1_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mperblock_nblock_nperblock,
c1_grid_desc_mblock_mperblock_nblock_nperblock,
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
block_2_ctile_map,
c0_matrix_mask);
#else
......@@ -99,15 +112,17 @@ __global__ void
ignore = p_b_grid;
ignore = p_b1_grid;
ignore = p_c_grid;
ignore = p_d0s_grid;
ignore = a_element_op;
ignore = b_element_op;
ignore = acc_element_op;
ignore = c0de_element_op;
ignore = b1_element_op;
ignore = c_element_op;
ignore = c1de_element_op;
ignore = a_grid_desc_ak0_m_ak1;
ignore = b_grid_desc_bk0_n_bk1;
ignore = b1_grid_desc_bk0_n_bk1;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = c1_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5;
ignore = block_2_ctile_map;
ignore = batch_count;
ignore = compute_base_ptr_of_batch;
......@@ -127,15 +142,15 @@ template <index_t NumDimG,
typename BDataType,
typename B1DataType,
typename CDataType,
typename Acc0BiasDataType,
typename Acc1BiasDataType,
typename D0sDataType,
typename D1sDataType,
typename GemmAccDataType,
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename AccElementwiseOperation,
typename C0DEElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation,
typename C1DEElementwiseOperation,
GemmSpecialization GemmSpec,
TensorSpecialization ASpec,
TensorSpecialization BSpec,
......@@ -193,23 +208,23 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
BDataType,
B1DataType,
CDataType,
Acc0BiasDataType,
Acc1BiasDataType,
D0sDataType,
D1sDataType,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
C0DEElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation,
C1DEElementwiseOperation,
MaskingSpec>
{
static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
"Number of dimension must be greater than 0");
static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
static constexpr index_t NumD0Tensor = D0sDataType::Size();
static constexpr index_t NumD1Tensor = D1sDataType::Size();
// TODO ANT: implement bias combination
static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
static_assert(NumD1Tensor == 0, "Gemm1 Bias addition is unimplemented");
#if 0
// TODO ANT: use alias
......@@ -262,14 +277,40 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
Number<B1K1>{});
}
static auto MakeD0sGridDescriptor_M_N(
const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_strides)
{
return generate_tuple(
[&](auto i) {
return Transform::MakeCGridDescriptor_M_N(acc0_biases_gs_ms_ns_lengths[i],
acc0_biases_gs_ms_ns_strides[i]);
},
Number<NumD0Tensor>{});
}
static auto MakeD0sGridDescriptor_G_M_N(
const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_strides)
{
return generate_tuple(
[&](auto i) {
return Transform::MakeCGridDescriptor_G_M_N(acc0_biases_gs_ms_ns_lengths[i],
acc0_biases_gs_ms_ns_strides[i]);
},
Number<NumD0Tensor>{});
}
using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
using CGridDesc_M_N = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
using C1GridDesc_M_N = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
using AGridDesc_G_M_K = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
using BGridDesc_G_N_K = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
using B1GridDesc_G_N_K = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
using CGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
using C1GridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
using D0sGridDesc_M_N = decltype(MakeD0sGridDescriptor_M_N({}, {}));
using D0sGridDesc_G_M_N = decltype(MakeD0sGridDescriptor_G_M_N({}, {}));
constexpr static auto make_MaskOutPredicate()
{
......@@ -289,11 +330,13 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
const BGridDesc_G_N_K& b_grid_desc_g_n_k,
const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
const CGridDesc_G_M_N& c_grid_desc_g_m_n)
const C1GridDesc_G_M_N& c1_grid_desc_g_m_n,
const D0sGridDesc_G_M_N& d0s_grid_desc_g_m_n)
: a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
c_grid_desc_g_m_n_(c_grid_desc_g_m_n)
c1_grid_desc_g_m_n_(c1_grid_desc_g_m_n),
d0s_grid_desc_g_m_n_(d0s_grid_desc_g_m_n)
{
}
......@@ -314,32 +357,42 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
__host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
{
return c_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
return c1_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
template <index_t I>
__host__ __device__ constexpr long_index_t GetD0BasePtr(index_t g_idx,
Number<I> d0_idx) const
{
return d0s_grid_desc_g_m_n_[d0_idx].CalculateOffset(make_multi_index(g_idx, 0, 0));
}
private:
AGridDesc_G_M_K a_grid_desc_g_m_k_;
BGridDesc_G_N_K b_grid_desc_g_n_k_;
B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_;
C1GridDesc_G_M_N c1_grid_desc_g_m_n_;
D0sGridDesc_G_M_N d0s_grid_desc_g_m_n_;
};
// GridwiseGemm
using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
using GridwiseGemm = GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle<
ADataType, // TODO: distinguish A/B datatype
GemmAccDataType,
CShuffleDataType,
CDataType,
D0sDataType,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
C0DEElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation,
C1DEElementwiseOperation,
InMemoryDataOperationEnum::Set,
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
B1GridDesc_BK0_N_BK1,
CGridDesc_M_N,
C1GridDesc_M_N,
D0sGridDesc_M_N,
NumGemmKPrefetchStage,
BlockSize,
MPerBlock,
......@@ -396,8 +449,8 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
const BDataType* p_b_grid,
const B1DataType* p_b1_grid,
CDataType* p_c_grid,
const std::array<void*, NumAcc0Bias> p_acc0_biases,
const std::array<void*, NumAcc1Bias> p_acc1_biases,
const std::array<void*, NumD0Tensor> p_acc0_biases,
const std::array<void*, NumD1Tensor> p_acc1_biases,
const std::vector<index_t>& a_gs_ms_ks_lengths,
const std::vector<index_t>& a_gs_ms_ks_strides,
const std::vector<index_t>& b_gs_ns_ks_lengths,
......@@ -406,44 +459,48 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
const std::vector<index_t>& c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
const std::vector<index_t>& c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
const std::array<std::vector<ck::index_t>, NumAcc1Bias>
const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_strides,
const std::array<std::vector<ck::index_t>, NumD1Tensor>&
acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
const std::array<std::vector<ck::index_t>, NumAcc1Bias>
const std::array<std::vector<ck::index_t>, NumD1Tensor>&
acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
AccElementwiseOperation acc_element_op,
C0DEElementwiseOperation c0de_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op)
C1DEElementwiseOperation c1de_element_op)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_b1_grid_{p_b1_grid},
p_c_grid_{p_c_grid},
p_d0s_grid_{},
a_grid_desc_ak0_m_ak1_{
DeviceOp::MakeAGridDescriptor_AK0_M_AK1(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
b_grid_desc_bk0_n_bk1_{
DeviceOp::MakeBGridDescriptor_BK0_N_BK1(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
b1_grid_desc_bk0_n_bk1_{DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(
b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
c_grid_desc_m_n_{Transform::MakeCGridDescriptor_M_N(c_gs_ms_gemm1ns_lengths,
c_gs_ms_gemm1ns_strides)},
c1_grid_desc_m_n_{Transform::MakeCGridDescriptor_M_N(c_gs_ms_gemm1ns_lengths,
c_gs_ms_gemm1ns_strides)},
a_grid_desc_g_m_k_{
Transform::MakeAGridDescriptor_G_M_K(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
b_grid_desc_g_n_k_{
Transform::MakeB0GridDescriptor_G_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
b1_grid_desc_g_n_k_{Transform::MakeB1GridDescriptor_G_N_K(
b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
c_grid_desc_g_m_n_{Transform::MakeCGridDescriptor_G_M_N(c_gs_ms_gemm1ns_lengths,
c_gs_ms_gemm1ns_strides)},
c_grid_desc_mblock_mperblock_nblock_nperblock_{},
block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
c1_grid_desc_g_m_n_{Transform::MakeCGridDescriptor_G_M_N(c_gs_ms_gemm1ns_lengths,
c_gs_ms_gemm1ns_strides)},
d0s_grid_desc_g_m_n_{DeviceOp::MakeD0sGridDescriptor_G_M_N(
acc0_biases_gs_ms_ns_lengths, acc0_biases_gs_ms_ns_strides)},
c1_grid_desc_mblock_mperblock_nblock_nperblock_{},
d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_{},
block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c1_grid_desc_m_n_)},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
acc_element_op_{acc_element_op},
c0de_element_op_{c0de_element_op},
b1_element_op_{b1_element_op},
c_element_op_{c_element_op},
c1de_element_op_{c1de_element_op},
c0_matrix_mask_{b_grid_desc_g_n_k_.GetLength(I1)},
raw_lengths_mz_nz_kz_gemm1nz_{a_gs_ms_ks_lengths[NumDimG + NumDimM - 1],
b_gs_ns_ks_lengths[NumDimG + NumDimN - 1],
......@@ -457,27 +514,39 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO + NumDimN - 1]},
c_mz_gemm1nz_strides_{c_gs_ms_gemm1ns_strides[NumDimG + NumDimM - 1],
c_gs_ms_gemm1ns_strides[NumDimG + NumDimM + NumDimO - 1]},
batch_count_{c_grid_desc_g_m_n_.GetLength(I0)},
compute_base_ptr_of_batch_{
a_grid_desc_g_m_k_, b_grid_desc_g_n_k_, b1_grid_desc_g_n_k_, c_grid_desc_g_m_n_}
batch_count_{c1_grid_desc_g_m_n_.GetLength(I0)},
compute_base_ptr_of_batch_{a_grid_desc_g_m_k_,
b_grid_desc_g_n_k_,
b1_grid_desc_g_n_k_,
c1_grid_desc_g_m_n_,
d0s_grid_desc_g_m_n_}
{
// TODO ANT: implement bias addition
ignore = p_acc0_biases;
ignore = p_acc1_biases;
ignore = acc0_biases_gs_ms_ns_lengths;
ignore = acc0_biases_gs_ms_ns_strides;
ignore = acc1_biases_gs_ms_gemm1ns_lengths;
ignore = acc1_biases_gs_ms_gemm1ns_strides;
static_for<0, NumD0Tensor, 1>{}([&](auto i) {
using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
// D0 pointer
p_d0s_grid_(i) = static_cast<const D0DataType*>(p_acc0_biases[i]);
});
if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
b_grid_desc_bk0_n_bk1_,
b1_grid_desc_bk0_n_bk1_,
c_grid_desc_m_n_,
c1_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n_);
c1_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c1_grid_desc_m_n_);
D0sGridDesc_M_N d0s_grid_desc_m_n{DeviceOp::MakeD0sGridDescriptor_M_N(
acc0_biases_gs_ms_ns_lengths, acc0_biases_gs_ms_ns_strides)};
d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_ =
GridwiseGemm::MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
d0s_grid_desc_m_n);
}
}
......@@ -492,9 +561,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
std::cout << "b1_grid_desc_g_n_k_: " << b1_grid_desc_g_n_k_.GetLength(I0) << ", "
<< b1_grid_desc_g_n_k_.GetLength(I1) << ", "
<< b1_grid_desc_g_n_k_.GetLength(I2) << '\n';
std::cout << "c_grid_desc_g_m_n_: " << c_grid_desc_g_m_n_.GetLength(I0) << ", "
<< c_grid_desc_g_m_n_.GetLength(I1) << ", "
<< c_grid_desc_g_m_n_.GetLength(I2) << '\n';
std::cout << "c1_grid_desc_g_m_n_: " << c1_grid_desc_g_m_n_.GetLength(I0) << ", "
<< c1_grid_desc_g_m_n_.GetLength(I1) << ", "
<< c1_grid_desc_g_m_n_.GetLength(I2) << '\n';
}
// pointers
......@@ -502,18 +571,23 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
const BDataType* p_b_grid_;
const B1DataType* p_b1_grid_;
CDataType* p_c_grid_;
typename GridwiseGemm::D0sGridPointer p_d0s_grid_;
// tensor descriptor
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
CGridDesc_M_N c_grid_desc_m_n_;
C1GridDesc_M_N c1_grid_desc_m_n_;
AGridDesc_G_M_K a_grid_desc_g_m_k_;
BGridDesc_G_N_K b_grid_desc_g_n_k_;
B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_;
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock_;
C1GridDesc_G_M_N c1_grid_desc_g_m_n_;
D0sGridDesc_G_M_N d0s_grid_desc_g_m_n_;
typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c1_grid_desc_mblock_mperblock_nblock_nperblock_;
typename GridwiseGemm::D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_;
// block-to-c-tile map
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
......@@ -521,9 +595,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
// element-wise op
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
AccElementwiseOperation acc_element_op_;
C0DEElementwiseOperation c0de_element_op_;
B1ElementwiseOperation b1_element_op_;
CElementwiseOperation c_element_op_;
C1DEElementwiseOperation c1de_element_op_;
// check C0 masking and padding
C0MatrixMask c0_matrix_mask_;
......@@ -552,7 +626,7 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
}
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.batch_count_;
arg.block_2_ctile_map_.CalculateGridSize(arg.c1_grid_desc_m_n_) * arg.batch_count_;
// Gemm0_K
const auto K =
......@@ -565,15 +639,17 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
typename GridwiseGemm::D0sGridPointer,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
C0DEElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation,
C1DEElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
DeviceOp::B1GridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename GridwiseGemm::D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
typename GridwiseGemm::DefaultBlock2CTileMap,
ComputeBasePtrOfStridedBatch,
C0MatrixMask,
......@@ -588,15 +664,17 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
arg.p_b_grid_,
arg.p_b1_grid_,
arg.p_c_grid_,
arg.p_d0s_grid_,
arg.a_element_op_,
arg.b_element_op_,
arg.acc_element_op_,
arg.c0de_element_op_,
arg.b1_element_op_,
arg.c_element_op_,
arg.c1de_element_op_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.b1_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.c1_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
arg.block_2_ctile_map_,
arg.batch_count_,
arg.compute_base_ptr_of_batch_,
......@@ -646,9 +724,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
// TODO ANT: Check if tensor specialization & strides mismatch
// Check if C permute dimension matches GEMM + GEMM shape
const index_t c_g = arg.c_grid_desc_g_m_n_.GetLength(I0); // unpadded
const index_t c_m = arg.c_grid_desc_m_n_.GetLength(I0);
const index_t c_gemm1n = arg.c_grid_desc_m_n_.GetLength(I1);
const index_t c_g = arg.c1_grid_desc_g_m_n_.GetLength(I0); // unpadded
const index_t c_m = arg.c1_grid_desc_m_n_.GetLength(I0);
const index_t c_gemm1n = arg.c1_grid_desc_m_n_.GetLength(I1);
const index_t a_m = arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
const index_t b1_gemm1n = arg.b1_grid_desc_bk0_n_bk1_.GetLength(I1);
......@@ -698,7 +776,7 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.b1_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_m_n_,
arg.c1_grid_desc_m_n_,
arg.block_2_ctile_map_);
}
......@@ -713,8 +791,8 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
const BDataType* p_b,
const B1DataType* p_b1,
CDataType* p_c,
const std::array<void*, NumAcc0Bias> p_acc0_biases,
const std::array<void*, NumAcc1Bias> p_acc1_biases,
const std::array<void*, NumD0Tensor> p_acc0_biases,
const std::array<void*, NumD1Tensor> p_acc1_biases,
const std::vector<index_t>& a_gs_ms_ks_lengths,
const std::vector<index_t>& a_gs_ms_ks_strides,
const std::vector<index_t>& b_gs_ns_ks_lengths,
......@@ -723,17 +801,17 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
const std::vector<index_t>& c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
const std::vector<index_t>& c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
const std::array<std::vector<ck::index_t>, NumAcc1Bias>
const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_strides,
const std::array<std::vector<ck::index_t>, NumD1Tensor>
acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
const std::array<std::vector<ck::index_t>, NumAcc1Bias>
const std::array<std::vector<ck::index_t>, NumD1Tensor>
acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
AccElementwiseOperation acc_element_op,
C0DEElementwiseOperation c0de_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op)
C1DEElementwiseOperation c1de_element_op)
{
return Argument{p_a,
p_b,
......@@ -755,9 +833,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
a_element_op,
b_element_op,
acc_element_op,
c0de_element_op,
b1_element_op,
c_element_op};
c1de_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
......@@ -769,8 +847,8 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
const void* p_b,
const void* p_b1,
void* p_c,
const std::array<void*, NumAcc0Bias> p_acc0_biases,
const std::array<void*, NumAcc1Bias> p_acc1_biases,
const std::array<void*, NumD0Tensor> p_acc0_biases,
const std::array<void*, NumD1Tensor> p_acc1_biases,
const std::vector<index_t>& a_gs_ms_ks_lengths,
const std::vector<index_t>& a_gs_ms_ks_strides,
const std::vector<index_t>& b_gs_ns_ks_lengths,
......@@ -779,17 +857,17 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
const std::vector<index_t>& c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
const std::vector<index_t>& c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
const std::array<std::vector<ck::index_t>, NumAcc1Bias>
const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_strides,
const std::array<std::vector<ck::index_t>, NumD1Tensor>
acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
const std::array<std::vector<ck::index_t>, NumAcc1Bias>
const std::array<std::vector<ck::index_t>, NumD1Tensor>
acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
AccElementwiseOperation acc_element_op,
C0DEElementwiseOperation c0de_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op) override
C1DEElementwiseOperation c1de_element_op) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
......@@ -811,9 +889,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
acc1_biases_gs_ms_gemm1ns_strides,
a_element_op,
b_element_op,
acc_element_op,
c0de_element_op,
b1_element_op,
c_element_op);
c1de_element_op);
}
// polymorphic
......
include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp
View file @ f8a6c69c
......@@ -533,6 +533,11 @@ struct DeviceElementwiseNormalizationImpl
return (false);
}
if(p_arg_->x_lds_size_ >= 65536)
{
return (false);
}
return true;
};
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
View file @ f8a6c69c
......@@ -670,6 +670,9 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
{
throw std::runtime_error("wrong! GridwiseGemmWelford has invalid setting");
}
if(arg.p_workspace_e_grid_ == nullptr || arg.p_workspace_mean_ == nullptr ||
arg.p_workspace_var_ == nullptr || arg.p_workspace_count_ == nullptr)
throw std::runtime_error("wrong! WorkSpace pointer has not been set");
index_t grid_size = arg.block_2_etile_map_.CalculateGridSize(arg.gemm_e_grid_desc_m_n_);
......@@ -941,7 +944,11 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
}
}
return true;
return GridwiseGemmWelford::CheckValidity(arg.a_grid_desc_m_k_,
arg.b_grid_desc_n_k_,
arg.ds_grid_desc_m_n_,
arg.gemm_e_grid_desc_m_n_,
arg.block_2_etile_map_);
}
// polymorphic
......@@ -1057,7 +1064,12 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<< GemmKPerBlock << ", "
<< AK1 << ", "
<< BK1 << ", "
<< getGemmSpecializationString(GemmSpec)
<< getGemmSpecializationString(GemmSpec) << ", "
<< PostShuffleThreadClusterSize_M_N::At(I0) << ", "
<< PostShuffleThreadClusterSize_M_N::At(I1) << ", "
<< LayernormThreadClusterSize_M_N::At(I0) << ", "
<< LayernormThreadClusterSize_M_N::At(I1) << ", "
<< LayernormThreadSliceSize_M
<< ">"
<< " LoopScheduler: "
<< LoopSchedToString[LoopSched] << ", "
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp 0 → 100644
View file @ f8a6c69c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AccDataType,
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
GemmSpecialization GemmSpec,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t K0PerBlock,
ck::index_t K1,
ck::index_t MPerWMMA,
ck::index_t NPerWMMA,
ck::index_t MRepeat,
ck::index_t NRepeat,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
ck::index_t ABlockTransferSrcVectorDim,
ck::index_t ABlockTransferSrcScalarPerVector,
ck::index_t ABlockTransferDstScalarPerVector_K1,
bool ABlockLdsAddExtraM,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
ck::index_t BBlockTransferSrcVectorDim,
ck::index_t BBlockTransferSrcScalarPerVector,
ck::index_t BBlockTransferDstScalarPerVector_K1,
bool BBlockLdsAddExtraN,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
ck::index_t NumPrefetch = 1,
ck::LoopScheduler LoopSched = make_default_loop_scheduler(),
ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation>
{
using DeviceOp = DeviceGemmMultipleD_Wmma_CShuffle;
static constexpr index_t NumDTensor = DsDataType::Size();
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
// K1 = Max Vector Access Pixels
static constexpr auto K1Number = Number<K1>{};
static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
{
assert(K % K1 == 0);
const index_t K0 = K / K1;
const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
#ifdef ENABLE_COLMAJOR
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
#endif
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_right_pad_transform(M, PadM)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
}
static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
{
assert(K % K1 == 0);
const index_t K0 = K / K1;
const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
}
template <typename ELayout_>
static auto MakeEGridDescriptor_M_N(index_t M, index_t N, index_t StrideE)
{
const auto e_grid_desc_m_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout_>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideE, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout_>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideE));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
e_grid_desc_m_n,
make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
e_grid_desc_m_n,
make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
}
static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& Ms,
const std::array<index_t, NumDTensor>& Ns,
const std::array<index_t, NumDTensor>& DsStride)
{
return generate_tuple(
[&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(Ms[i], Ns[i], DsStride[i]);
},
Number<NumDTensor>{});
}
// Gridwise descriptor, mapping to whole given provblem.
using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
// GridwiseOp
using GridwiseOp = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle<
// DataType Family
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
// InMemory Data Descriptor
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
DsGridDesc_M_N,
EGridDesc_M_N,
// ElementwiseOp Family
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
InMemoryDataOperationEnum::Set,
// Tiling Family
MPerBlock,
NPerBlock,
K0PerBlock,
MPerWMMA,
NPerWMMA,
K1,
MRepeat,
NRepeat,
// ThreadCluster Family
BlockSize,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
false, // AThreadTransferSrcResetCoordinateAfterRun,
ABlockLdsAddExtraM,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
false, // BThreadTransferSrcResetCoordinateAfterRun,
BBlockLdsAddExtraN,
CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CDEShuffleBlockTransferScalarPerVector_NPerBlock,
NumPrefetch,
LoopSched,
PipelineVer>;
// Argument
struct Argument : public BaseArgument
{
Argument(const void* p_a_grid,
const void* p_b_grid,
std::array<const void*, NumDTensor> p_ds_grid,
void* p_e_grid,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
std::array<index_t, NumDTensor> StrideDs,
index_t StrideE,
index_t M01,
index_t N01,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
p_ds_grid_{},
p_e_grid_{static_cast<EDataType*>(p_e_grid)},
a_grid_desc_k0_m_k1_{},
b_grid_desc_k0_n_k1_{},
ds_grid_desc_m_n_{},
e_grid_desc_m_n_{},
ds_grid_desc_mblock_mperblock_nblock_nperblock{},
e_grid_desc_mblock_mperblock_nblock_nperblock{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
a_grid_desc_k0_m_k1_ = DeviceOp::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
b_grid_desc_k0_n_k1_ = DeviceOp::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
static_for<0, NumDTensor, 1>{}([&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
// D pointer
p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
// D desc
ds_grid_desc_m_n_(i) =
DeviceOp::MakeEGridDescriptor_M_N<DLayout>(M, N, StrideDs[i]);
});
e_grid_desc_m_n_ = DeviceOp::MakeEGridDescriptor_M_N<ELayout>(M, N, StrideE);
block_2_ctile_map_ = GridwiseOp::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_, M01, N01);
if(GridwiseOp::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
ds_grid_desc_m_n_,
e_grid_desc_m_n_,
block_2_ctile_map_))
{
ds_grid_desc_mblock_mperblock_nblock_nperblock =
GridwiseOp::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
ds_grid_desc_m_n_);
e_grid_desc_mblock_mperblock_nblock_nperblock =
GridwiseOp::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
e_grid_desc_m_n_);
}
}
// Pointers
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
typename GridwiseOp::DsGridPointer p_ds_grid_;
EDataType* p_e_grid_;
// Tensor Descriptors
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
DsGridDesc_M_N ds_grid_desc_m_n_;
EGridDesc_M_N e_grid_desc_m_n_;
typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock;
typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock;
// Block to Tile mapping
typename GridwiseOp::DefaultBlock2CTileMap block_2_ctile_map_;
// Idle
index_t M01_;
index_t N01_;
// ElementwiseOp
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
#if 0
{
std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
<< ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
<< arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
<< ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
<< arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0)
<< ", " << arg.c_grid_desc_m_n_.GetLength(I1) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I2) << "}" << std::endl;
}
#endif
if(!GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_m_n_,
arg.e_grid_desc_m_n_,
arg.block_2_ctile_map_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_k0mk1_k0nk1_m0nm1_wmma_v1r1 has invalid setting");
}
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
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(GridwiseOp::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_mupltipe_d_wmma_cshuffle<
GridwiseOp,
ADataType,
BDataType,
typename GridwiseOp::DsGridPointer,
EDataType,
remove_reference_t<typename DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<typename DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
remove_reference_t<
typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
remove_reference_t<typename GridwiseOp::DefaultBlock2CTileMap>,
true>; // Last Option is W/O
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_ds_grid_,
arg.p_e_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock,
arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.block_2_ctile_map_);
}
else
{
const auto kernel = kernel_gemm_mupltipe_d_wmma_cshuffle<
GridwiseOp,
ADataType,
BDataType,
typename GridwiseOp::DsGridPointer,
EDataType,
remove_reference_t<typename DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<typename DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
remove_reference_t<
typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
remove_reference_t<typename GridwiseOp::DefaultBlock2CTileMap>,
false>;
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_ds_grid_,
arg.p_e_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock,
arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.block_2_ctile_map_);
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx1100")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{
return false;
}
}
else
{
return false;
}
return GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_m_n_,
arg.e_grid_desc_m_n_,
arg.block_2_ctile_map_);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
std::array<ck::index_t, NumDTensor> StrideDs,
index_t StrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{p_a,
p_b,
p_ds,
p_e,
M,
N,
K,
StrideA,
StrideB,
StrideDs,
StrideE,
1,
1,
a_element_op,
b_element_op,
cde_element_op};
}
// polymorphic
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
std::array<ck::index_t, NumDTensor> StrideDs,
index_t StrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) override
{
return std::make_unique<Argument>(p_a,
p_b,
p_ds,
p_e,
M,
N,
K,
StrideA,
StrideB,
StrideDs,
StrideE,
1,
1,
a_element_op,
b_element_op,
cde_element_op);
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
std::map<LoopScheduler, std::string> LoopSchedToString{
{LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
{PipelineVersion::v2, "v2"}};
// clang-format off
str << "DeviceGemmMultipleD_Wmma_CShuffle"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< K1 << ", "
<< MPerWMMA << ", "
<< NPerWMMA << ", "
<< MRepeat << ", "
<< NRepeat
<< ">"
<< " NumPrefetch: "
<< NumPrefetch << ", "
<< "LoopScheduler: "
<< LoopSchedToString[LoopSched] << ", "
<< "PipelineVersion: "
<< PipelineVersionToString[PipelineVer];
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp 0 → 100644
View file @ f8a6c69c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <numeric>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace {
struct ComputePtrOffsetOfStridedBatch
{
__host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideA_);
}
__host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideB_);
}
__host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideC_);
}
index_t BatchStrideA_;
index_t BatchStrideB_;
index_t BatchStrideC_;
};
} // namespace
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename AGridDesc_B_K0_M0_M1_K1,
typename BGridDesc_B_K0_N0_N1_K1,
typename CGridDesc_M0_M10_M11_N0_N10_N11,
typename Block2CTileMap,
typename ComputePtrOffsetOfBatch,
bool HasMainKBlockLoop,
bool HasDoubleTailKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_batched_gemm_dlops_bwd_weight(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const index_t batch_count,
const AGridDesc_B_K0_M0_M1_K1 a_grid_desc_kbatch_k0_m0_m1_k1,
const BGridDesc_B_K0_N0_N1_K1 b_grid_desc_kbatch_k0_n0_n1_k1,
const CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11,
const Block2CTileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
__shared__ FloatAB p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB)];
GridwiseGemm::template Run<HasMainKBlockLoop, HasDoubleTailKBlockLoop>(
p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset,
p_c_grid + c_batch_offset,
p_shared,
a_grid_desc_kbatch_k0_m0_m1_k1,
b_grid_desc_kbatch_k0_n0_n1_k1,
c_grid_desc_m0_m10_m11_n0_n10_n11,
block_2_ctile_map,
integral_constant<bool, HasMainKBlockLoop>{},
integral_constant<bool, HasDoubleTailKBlockLoop>{});
}
template <ck::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType,
typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation,
ConvolutionBackwardWeightSpecialization ConvBackwardWeightSpecialization,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t K0PerBlock,
ck::index_t K1,
index_t M1PerThread,
index_t N1PerThread,
index_t KPerThread,
typename M1N1ThreadClusterM1Xs,
typename M1N1ThreadClusterN1Xs,
typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
typename ABlockTransferSrcVectorTensorContiguousDimOrder,
typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
typename BBlockTransferSrcVectorTensorContiguousDimOrder,
typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
typename CThreadTransferSrcDstAccessOrder,
index_t CThreadTransferSrcDstVectorDim,
index_t CThreadTransferDstScalarPerVector>
struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
: public DeviceGroupedConvBwdWeight<
NDimSpatial,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWC,
ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GNDHWC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GKXC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::GKZYXC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWK,
ck::tensor_layout::convolution::GNHWK,
ck::tensor_layout::convolution::GNDHWK>>,
InDataType,
WeiDataType,
OutDataType,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation>
{
using DeviceOp = DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl;
using ADataType = OutDataType;
using BDataType = InDataType;
using CDataType = WeiDataType;
using AElementwiseOperation = OutElementwiseOperation;
using BElementwiseOperation = InElementwiseOperation;
using CElementwiseOperation = WeiElementwiseOperation;
// TODO make A/B datatype different
using ABDataType = InDataType;
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>{};
static constexpr auto I5 = Number<5>{};
static constexpr auto K1Number = Number<K1>{};
static constexpr auto GemmK1Number = 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;
template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
std::array<ck::index_t, NDimSpatial> input_left_pads,
std::array<ck::index_t, NDimSpatial> input_right_pads,
ck::index_t batch_k)
{
using namespace ck;
const index_t Wi = input_spatial_lengths[0];
const index_t Wo = output_spatial_lengths[0];
const index_t X = filter_spatial_lengths[0];
const index_t InLeftPadW = input_left_pads[0];
const index_t InRightPadW = input_right_pads[0];
const index_t ConvStrideW = conv_filter_strides[0];
const index_t ConvDilationW = conv_filter_dilations[0];
const index_t GemmKTotal = N * Wo;
const index_t GemmM = K;
const index_t GemmN = C * X;
const index_t GemmKBatch = batch_k;
const index_t GemmK0 =
math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
K0PerBlock;
const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
// A: output tensor
const auto out_gemmktotal_gemmm_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K));
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 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>{}));
// B: input tensor
const auto in_gemmktotal_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N * Wi, C));
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(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
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(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// C: weights tensor
const auto wei_gemmm_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, X * C));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
}
else
{
const auto out_gemmktotal_gemmm_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K));
const auto in_n_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, 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 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>{}));
// B: input tensor
const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
in_n_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
const auto in_gemmktotal_gemmn_grid_desc =
transform_tensor_descriptor(in_n_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(X, C)),
make_merge_transform(make_tuple(N, Wo))),
make_tuple(Sequence<1, 3>{}, Sequence<0, 2>{}),
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>{}));
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>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, X * C));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
}
} // function end
template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
std::array<ck::index_t, NDimSpatial> input_left_pads,
std::array<ck::index_t, NDimSpatial> input_right_pads,
ck::index_t batch_k)
{
using namespace ck;
const index_t Hi = input_spatial_lengths[0];
const index_t Wi = input_spatial_lengths[1];
const index_t Ho = output_spatial_lengths[0];
const index_t Wo = output_spatial_lengths[1];
const index_t Y = filter_spatial_lengths[0];
const index_t X = filter_spatial_lengths[1];
const index_t InLeftPadH = input_left_pads[0];
const index_t InLeftPadW = input_left_pads[1];
const index_t InRightPadH = input_right_pads[0];
const index_t InRightPadW = input_right_pads[1];
const index_t ConvStrideH = conv_filter_strides[0];
const index_t ConvStrideW = conv_filter_strides[1];
const index_t ConvDilationH = conv_filter_dilations[0];
const index_t ConvDilationW = conv_filter_dilations[1];
const index_t GemmKTotal = N * Ho * Wo;
const index_t GemmM = K;
const index_t GemmN = C * X * Y;
const index_t GemmKBatch = batch_k;
const index_t GemmK0 =
math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
K0PerBlock;
const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
// A: output tensor
const auto out_gemmktotal_gemmm_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
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 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>{}));
// B: input tensor
const auto in_gemmktotal_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N * Hi * Wi, C));
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(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
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(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
}
else
{
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 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>{}));
// B: input tensor
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
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 =
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>{}));
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>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
}
} // function end
template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
std::array<ck::index_t, NDimSpatial> input_left_pads,
std::array<ck::index_t, NDimSpatial> input_right_pads,
ck::index_t batch_k)
{
using namespace ck;
const index_t Di = input_spatial_lengths[0];
const index_t Hi = input_spatial_lengths[1];
const index_t Wi = input_spatial_lengths[2];
const index_t Do = output_spatial_lengths[0];
const index_t Ho = output_spatial_lengths[1];
const index_t Wo = output_spatial_lengths[2];
const index_t Z = filter_spatial_lengths[0];
const index_t Y = filter_spatial_lengths[1];
const index_t X = filter_spatial_lengths[2];
const index_t InLeftPadD = input_left_pads[0];
const index_t InLeftPadH = input_left_pads[1];
const index_t InLeftPadW = input_left_pads[2];
const index_t InRightPadD = input_right_pads[0];
const index_t InRightPadH = input_right_pads[1];
const index_t InRightPadW = input_right_pads[2];
const index_t ConvStrideD = conv_filter_strides[0];
const index_t ConvStrideH = conv_filter_strides[1];
const index_t ConvStrideW = conv_filter_strides[2];
const index_t ConvDilationD = conv_filter_dilations[0];
const index_t ConvDilationH = conv_filter_dilations[1];
const index_t ConvDilationW = conv_filter_dilations[2];
const index_t GemmKTotal = N * Do * Ho * Wo;
const index_t GemmM = K;
const index_t GemmN = C * Z * X * Y;
const index_t GemmKBatch = batch_k;
const index_t GemmK0 =
math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
K0PerBlock;
const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
// A: output tensor
const auto out_gemmktotal_gemmm_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
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 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>{}));
// B: input tensor
const auto in_gemmktotal_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N * Di * Hi * Wi, C));
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(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
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(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
}
else
{
const auto out_gemmktotal_gemmm_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
const auto in_n_di_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Di, 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 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>{}));
// B: input tensor
const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_di_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Di, InLeftPadD, InRightPadD),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_dip_hip_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(Sequence<0>{},
Sequence<1, 2>{},
Sequence<3, 4>{},
Sequence<5, 6>{},
Sequence<7>{}));
const auto in_gemmktotal_gemmn_grid_desc = transform_tensor_descriptor(
in_n_z_do_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(Z, Y, X, C)),
make_merge_transform(make_tuple(N, Do, Ho, Wo))),
make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}),
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>{}));
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>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
}
} // function end
template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<1>(
1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1}, 1);
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<2>(
1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, 1);
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<3>(1,
1,
1,
{1, 1, 1},
{1, 1, 1},
{1, 1, 1},
{1, 1, 1},
{1, 1, 1},
{1, 1, 1},
{1, 1, 1},
1);
}
using ABCGridDescs = decltype(GetABCGridDesc<NDimSpatial>());
using AGridDesc_B_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
using BGridDesc_B_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
using CGridDesc_M_N = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
using GridwiseGemm =
GridwiseGemmDl_bkm_bkn_mn_v1r3<BlockSize,
ADataType,
AccDataType,
CDataType,
InMemoryDataOperationEnum::Set,
AGridDesc_B_K0_M_K1,
BGridDesc_B_K0_N_K1,
CGridDesc_M_N,
MPerBlock,
NPerBlock,
K0PerBlock,
K1,
M1PerThread,
N1PerThread,
KPerThread,
M1N1ThreadClusterM1Xs,
M1N1ThreadClusterN1Xs,
ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
ABlockTransferSrcVectorTensorContiguousDimOrder,
ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
BBlockTransferSrcVectorTensorContiguousDimOrder,
BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector>;
// Argument
using AGridDesc_B_K0_M0_M1_K1 =
decltype(GridwiseGemm::MakeAGridDescriptor_B_K0_M0_M1_K1(AGridDesc_B_K0_M_K1{}));
using BGridDesc_B_K0_N0_N1_K1 =
decltype(GridwiseGemm::MakeBGridDescriptor_B_K0_N0_N1_K1(BGridDesc_B_K0_N_K1{}));
using CGridDesc_M0_M10_M11_N0_N10_N11 =
decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
using Block2CTileMap =
decltype(GridwiseGemm::MakeCBlockClusterAdaptor(CGridDesc_M_N{}, 1, 1, 1));
struct Argument : public BaseArgument
{
Argument(const InDataType* p_in_grid,
WeiDataType* p_wei_grid,
const OutDataType* p_out_grid,
ck::index_t G,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
std::array<ck::index_t, NDimSpatial> input_left_pads,
std::array<ck::index_t, NDimSpatial> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
ck::index_t split_k)
: p_a_grid_{p_out_grid},
p_b_grid_{p_in_grid},
p_c_grid_{p_wei_grid},
a_grid_desc_kbatch_k0_m_k1_{},
b_grid_desc_kbatch_k0_n_k1_{},
c_grid_desc_m_n_{},
block_2_ctile_map_{},
compute_ptr_offset_of_batch_{},
a_element_op_{out_element_op},
b_element_op_{wei_element_op},
c_element_op_{in_element_op},
Conv_G_{G},
Conv_N_{N},
Conv_K_{K},
Conv_C_{C},
input_spatial_lengths_{input_spatial_lengths},
filter_spatial_lengths_{filter_spatial_lengths},
output_spatial_lengths_{output_spatial_lengths},
conv_filter_strides_{conv_filter_strides},
conv_filter_dilations_{conv_filter_dilations},
input_left_pads_{input_left_pads},
input_right_pads_{input_right_pads},
k_batch_{split_k}
{
const auto descs =
DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
k_batch_);
a_grid_desc_kbatch_k0_m_k1_ = descs[I0];
b_grid_desc_kbatch_k0_n_k1_ = descs[I1];
c_grid_desc_m_n_ = descs[I2];
a_grid_desc_kbatch_k0_m0_m1_k1_ =
GridwiseGemm::MakeAGridDescriptor_B_K0_M0_M1_K1(a_grid_desc_kbatch_k0_m_k1_);
b_grid_desc_kbatch_k0_n0_n1_k1_ =
GridwiseGemm::MakeBGridDescriptor_B_K0_N0_N1_K1(b_grid_desc_kbatch_k0_n_k1_);
c_grid_desc_m0_m10_m11_n0_n10_n11_ =
GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(c_grid_desc_m_n_);
ck::index_t M01 = 1;
ck::index_t N01 = 1;
block_2_ctile_map_ =
GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
// A/B/C Batch Stride
compute_ptr_offset_of_batch_.BatchStrideA_ =
N * K *
std::accumulate(begin(output_spatial_lengths),
end(output_spatial_lengths),
index_t{1},
std::multiplies<>{});
compute_ptr_offset_of_batch_.BatchStrideB_ =
N * C *
std::accumulate(begin(input_spatial_lengths),
end(input_spatial_lengths),
index_t{1},
std::multiplies<>{});
compute_ptr_offset_of_batch_.BatchStrideC_ =
K * C *
std::accumulate(begin(filter_spatial_lengths),
end(filter_spatial_lengths),
index_t{1},
std::multiplies<>{});
}
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
AGridDesc_B_K0_M_K1 a_grid_desc_kbatch_k0_m_k1_;
BGridDesc_B_K0_N_K1 b_grid_desc_kbatch_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
AGridDesc_B_K0_M0_M1_K1 a_grid_desc_kbatch_k0_m0_m1_k1_;
BGridDesc_B_K0_N0_N1_K1 b_grid_desc_kbatch_k0_n0_n1_k1_;
CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11_;
// DefaultBlock2CTileMap block_2_ctile_map_;
Block2CTileMap block_2_ctile_map_;
// for computing batch offset
ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
// element-wise op
OutElementwiseOperation a_element_op_;
WeiElementwiseOperation b_element_op_;
InElementwiseOperation c_element_op_;
// for checking IsSupportedArgument()
index_t Conv_G_;
index_t Conv_N_;
index_t Conv_K_;
index_t Conv_C_;
std::array<ck::index_t, NDimSpatial> input_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> output_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> conv_filter_strides_;
std::array<ck::index_t, NDimSpatial> conv_filter_dilations_;
std::array<ck::index_t, NDimSpatial> input_left_pads_;
std::array<ck::index_t, NDimSpatial> input_right_pads_;
index_t k_batch_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
void ShowInfo(const Argument& arg)
{
std::cout << "arg.a_grid_desc_kbatch_k0_m_k1_{"
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0) << ", "
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1) << ", "
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I2) << ", "
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I3) << "}" << std::endl;
std::cout << "arg.b_grid_desc_kbatch_k0_n_k1_{"
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I0) << ", "
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I1) << ", "
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I2) << ", "
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I3) << "}" << std::endl;
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
}
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_))
{
throw std::runtime_error(
"wrong! GridwiseGemm GridwiseGemmDl_bkm_bkn_mn_v1r3 has invalid setting");
}
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.Conv_G_;
auto launch_kernel = [&](auto has_main_k_block_loop,
auto has_double_tail_k_block_loop) {
constexpr bool has_main_loop = has_main_k_block_loop.value;
constexpr bool has_double_loop = has_double_tail_k_block_loop.value;
const auto kernel = kernel_batched_gemm_dlops_bwd_weight<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceOp::AGridDesc_B_K0_M0_M1_K1>,
remove_reference_t<DeviceOp::BGridDesc_B_K0_N0_N1_K1>,
remove_reference_t<DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11>,
remove_reference_t<DeviceOp::Block2CTileMap>,
ComputePtrOffsetOfStridedBatch,
has_main_loop,
has_double_loop>;
return launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.Conv_G_,
arg.a_grid_desc_kbatch_k0_m0_m1_k1_,
arg.b_grid_desc_kbatch_k0_n0_n1_k1_,
arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
arg.block_2_ctile_map_,
arg.compute_ptr_offset_of_batch_);
};
const auto K0 = arg.a_grid_desc_kbatch_k0_m0_m1_k1_.GetLength(I1);
const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
const bool has_double_tail_k_block_loop =
GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
if(has_main_k_block_loop && has_double_tail_k_block_loop)
{
return launch_kernel(integral_constant<bool, true>{},
integral_constant<bool, true>{});
}
else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
{
return launch_kernel(integral_constant<bool, true>{},
integral_constant<bool, false>{});
}
else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
{
return launch_kernel(integral_constant<bool, false>{},
integral_constant<bool, true>{});
}
else
{
return launch_kernel(integral_constant<bool, false>{},
integral_constant<bool, false>{});
}
}
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
// check device
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
{
return false;
}
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
// check if it's 1x1, stride=1 pad = 0 conv
for(int i = 0; i < NDimSpatial; i++)
{
if(!(arg.filter_spatial_lengths_[i] == 1 && arg.conv_filter_strides_[i] == 1 &&
arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
{
return false;
}
}
}
// matrix A
{
auto srcVectorLengths = ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1{};
if(srcVectorLengths[I2] != 1 || srcVectorLengths[I3] != 1)
{
return false;
}
if(K1 % srcVectorLengths[I4] != 0 || K0PerBlock % srcVectorLengths[I1] != 0)
{
return false;
}
const index_t K = arg.Conv_K_;
if(K % (srcVectorLengths[I1] * srcVectorLengths[I4]) != 0)
{
return false;
}
}
// matrix B
{
auto srcLoadLenghts = BBlockTransferThreadSliceLengths_K0_N0_N1_K1{};
auto srcVectorLengths = BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1{};
if(srcVectorLengths[I1] != 1 || srcVectorLengths[I4] != 1)
{
return false;
}
if(srcLoadLenghts[I2] % srcVectorLengths[I2] != 0 ||
srcLoadLenghts[I3] % srcVectorLengths[I3] != 0)
{
return false;
}
const index_t C = arg.Conv_K_;
if(C % (srcVectorLengths[I2] * srcVectorLengths[I3]) != 0)
{
return false;
}
}
// vector store C matrix into global memory
if(!(arg.Conv_C_ % CThreadTransferDstScalarPerVector == 0))
{
std::cout << "Not surpport,because: arg.Conv_C_ % CThreadTransferDstScalarPerVector = "
<< arg.Conv_C_ % CThreadTransferDstScalarPerVector << std::endl;
return false;
}
// Gridwise GEMM size
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_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const InDataType* p_in_grid,
WeiDataType* p_wei_grid,
const OutDataType* p_out_grid,
ck::index_t G,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
std::array<ck::index_t, NDimSpatial> input_left_pads,
std::array<ck::index_t, NDimSpatial> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
ck::index_t split_k)
{
return Argument{p_in_grid,
p_wei_grid,
p_out_grid,
G,
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op,
split_k};
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_in_grid,
void* p_wei_grid,
const void* p_out_grid,
ck::index_t G,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
std::array<ck::index_t, NDimSpatial> input_left_pads,
std::array<ck::index_t, NDimSpatial> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
ck::index_t split_k) override
{
return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
static_cast<WeiDataType*>(p_wei_grid),
static_cast<const OutDataType*>(p_out_grid),
G,
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< getConvBackwardWeightSpecializationString(ConvBackwardWeightSpecialization)
<< ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp 0 → 100644
View file @ f8a6c69c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <functional>
#include <iostream>
#include <iterator>
#include <numeric>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/host_utility/io.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace {
template <index_t NumDTensor>
struct ComputePtrOffsetOfStridedBatch
{
ComputePtrOffsetOfStridedBatch() = default;
ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
index_t BatchStrideB,
Array<ck::index_t, NumDTensor> BatchStrideDs,
index_t BatchStrideE)
: BatchStrideA_(BatchStrideA),
BatchStrideB_(BatchStrideB),
BatchStrideDs_(BatchStrideDs),
BatchStrideE_(BatchStrideE)
{
}
__host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideA_);
}
__host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideB_);
}
__host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
{
Array<long_index_t, NumDTensor> ds_offset;
static_for<0, NumDTensor, 1>{}(
[&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
return ds_offset;
}
__host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideE_);
}
index_t BatchStrideA_;
index_t BatchStrideB_;
Array<ck::index_t, NumDTensor> BatchStrideDs_;
index_t BatchStrideE_;
};
} // namespace
//
// @brief Device Convolution operation.
//
// Supports:
// @li Forward convolution with up to 3 spatial dimentions
// @li Input tensor in GNWC data format
// @li Weight tensor in GKXC data format
// @li Output tensor in GNWK data format
//
// 1D:
// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
// 2D:
// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
// 3D:
// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
// Assume:
// AK1 == BK1
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AccDataType,
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
ConvolutionForwardSpecialization ConvForwardSpecialization,
GemmSpecialization GemmSpec,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t K0PerBlock,
ck::index_t K1,
ck::index_t MPerWMMA,
ck::index_t NPerWMMA,
ck::index_t MRepeat,
ck::index_t NRepeat,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
index_t NumGemmKPrefetchStage = 1,
LoopScheduler LoopSched = make_default_loop_scheduler(),
ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
: public DeviceGroupedConvFwdMultipleD<NDimSpatial,
ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation>
{
using DeviceOp = DeviceGroupedConvFwdMultipleD_Wmma_CShuffle;
static constexpr index_t NumDTensor = DsDataType::Size();
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 index_t KPerBlock = K0PerBlock * K1;
static constexpr auto conv_to_gemm_transformer =
TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
template <typename ALay>
static auto
MakeAGridDescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
const std::array<index_t, NDimSpatial>& input_right_pads)
{
const auto in_gemmmraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
e_g_n_k_wos_lengths,
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
const auto in_gemmm_gemmk_desc =
matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
return in_gemmm_gemmk_desc;
}
template <typename BLay>
static auto
MakeBGridDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
{
const auto wei_gemmnraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
b_g_k_c_xs_strides);
const auto wei_gemmn_gemmk_desc =
matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
return wei_gemmn_gemmk_desc;
}
template <typename ELay>
static auto
MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
{
const auto out_gemmmraw_gemmnraw_desc =
conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
e_g_n_k_wos_strides);
const auto out_gemmm_gemmn_desc =
matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
return out_gemmm_gemmn_desc;
}
static auto MakeDsGridDescriptor_M_N(
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides)
{
return generate_tuple(
[&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(ds_g_n_k_wos_lengths[i],
ds_g_n_k_wos_strides[i]);
},
Number<NumDTensor>{});
}
// desc for problem definition
using AGridDesc_M_K = remove_cvref_t<decltype(
MakeAGridDescriptor_M_K<ALayout>({}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
using BGridDesc_N_K = remove_cvref_t<decltype(MakeBGridDescriptor_N_K<BLayout>({}, {}))>;
using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
using EGridDesc_M_N = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
// A desc for source in blockwise copy
template <typename AGridDesc_M_K>
__host__ __device__ static constexpr auto
MakeAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
{
const auto M = a_grid_desc_m_k.GetLength(I0);
const auto K = a_grid_desc_m_k.GetLength(I1);
const auto AK1 = K1;
const auto AK0 = K / AK1;
return transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
// B desc for source in blockwise copy
template <typename BGridDesc_N_K>
__host__ __device__ static constexpr auto
MakeBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
{
const auto N = b_grid_desc_n_k.GetLength(I0);
const auto K = b_grid_desc_n_k.GetLength(I1);
const auto BK1 = K1;
const auto BK0 = K / BK1;
return transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
using AGridDesc_AK0_M_AK1 = decltype(DeviceOp::MakeAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}));
using BGridDesc_BK0_N_BK1 = decltype(DeviceOp::MakeBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}));
// GridwiseOp
using GridwiseOp = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle<
// DataType Family
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
// InMemory Data Descriptor
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
DsGridDesc_M_N,
EGridDesc_M_N,
// ElementwiseOp Family
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
InMemoryDataOperationEnum::Set,
// Tiling Family
MPerBlock,
NPerBlock,
K0PerBlock,
MPerWMMA,
NPerWMMA,
K1,
MRepeat,
NRepeat,
// ThreadCluster Family
BlockSize,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
false,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
false,
BBlockLdsExtraN,
CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CDEShuffleBlockTransferScalarPerVector_NPerBlock,
NumGemmKPrefetchStage,
LoopSched,
PipelineVer>;
// Argument
struct Argument : public BaseArgument
{
Argument(const void* p_a,
const void* p_b,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_lengths,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
const std::array<index_t, NDimSpatial>& input_right_pads,
index_t M01,
index_t N01,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op)
: p_a_grid_{static_cast<const ADataType*>(p_a)},
p_b_grid_{static_cast<const BDataType*>(p_b)},
p_ds_grid_{},
p_e_grid_{static_cast<EDataType*>(p_e)},
num_group_{a_g_n_c_wis_lengths[0]},
a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K<ALayout>(a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
e_g_n_k_wos_lengths,
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads)},
b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K<BLayout>(b_g_k_c_xs_lengths,
b_g_k_c_xs_strides)},
ds_grid_desc_m_n_{},
e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(e_g_n_k_wos_lengths,
e_g_n_k_wos_strides)},
a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
e_grid_desc_mblock_mperblock_nblock_nperblock_{},
block_2_etile_map_{GridwiseOp::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_, M01, N01)},
compute_ptr_offset_of_batch_{},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op},
a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
ds_g_n_k_wos_lengths_{ds_g_n_k_wos_lengths},
ds_g_n_k_wos_strides_{ds_g_n_k_wos_strides},
e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
conv_filter_strides_{conv_filter_strides},
conv_filter_dilations_{conv_filter_dilations},
input_left_pads_{input_left_pads},
input_right_pads_{input_right_pads}
{
// A/B/E Batch Stride
compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
// populate pointer, batch stride, desc for Ds
static_for<0, NumDTensor, 1>{}([&](auto i) {
// using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
// D pointer
p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
// D batch stride
compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
});
// D desc
ds_grid_desc_m_n_ =
DeviceOp::MakeDsGridDescriptor_M_N(ds_g_n_k_wos_lengths, ds_g_n_k_wos_strides);
// populate desc for Ds/E
e_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseOp::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n_);
ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseOp::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
ds_grid_desc_m_n_);
}
void Print() const
{
std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
static_for<0, NumDTensor, 1>{}(
[&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
}
// private:
// pointers
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
typename GridwiseOp::DsGridPointer p_ds_grid_;
EDataType* p_e_grid_;
// tensor descriptors for problem definiton
index_t num_group_;
AGridDesc_M_K a_grid_desc_m_k_;
BGridDesc_N_K b_grid_desc_n_k_;
DsGridDesc_M_N ds_grid_desc_m_n_;
EGridDesc_M_N e_grid_desc_m_n_;
// tensor descriptors for block/thread-wise copy
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock_;
typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock_;
// block-to-e-tile map
typename GridwiseOp::DefaultBlock2CTileMap block_2_etile_map_;
// for computing batch offset
ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
// element-wise op
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
// for checking IsSupportedArgument()
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
std::array<index_t, NDimSpatial> conv_filter_strides_;
std::array<index_t, NDimSpatial> conv_filter_dilations_;
std::array<index_t, NDimSpatial> input_left_pads_;
std::array<index_t, NDimSpatial> input_right_pads_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
if(stream_config.log_level_ > 0)
{
arg.Print();
}
const index_t grid_size =
arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.num_group_;
const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
auto launch_kernel = [&](auto has_main_k_block_loop) {
constexpr bool has_main_loop = has_main_k_block_loop.value;
const auto kernel = kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle<
GridwiseOp,
ADataType,
BDataType,
typename GridwiseOp::DsGridPointer,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
remove_reference_t<typename GridwiseOp::DefaultBlock2CTileMap>,
ComputePtrOffsetOfStridedBatch<NumDTensor>,
has_main_loop>;
return launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_ds_grid_,
arg.p_e_grid_,
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.a_g_n_c_wis_lengths_[0], // Group count
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.block_2_etile_map_,
arg.compute_ptr_offset_of_batch_);
};
if(GridwiseOp::CalculateHasMainKBlockLoop(K))
{
return launch_kernel(integral_constant<bool, true>{});
}
else
{
return launch_kernel(integral_constant<bool, false>{});
}
}
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static bool IsSupportedArgument(const Argument& arg)
{
namespace ctc = tensor_layout::convolution;
// check device
if(get_device_name() == "gfx1100")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{
return false;
}
}
else
{
return false;
}
// check ConvolutionForwardSpecialization
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
{
// check if it's 1x1, stride=1 conv
for(index_t i = 0; i < NDimSpatial; ++i)
{
const index_t X = arg.b_g_k_c_xs_lengths_[i + 2];
const index_t ConvStride = arg.conv_filter_strides_[i];
const index_t LeftPad = arg.input_left_pads_[i];
const index_t RightPad = arg.input_right_pads_[i];
if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
{
return false;
}
}
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization::Filter1x1Pad0)
{
// check if it's 1x1 conv
for(index_t i = 0; i < NDimSpatial; ++i)
{
const index_t X = arg.b_g_k_c_xs_lengths_[i + 2];
const index_t LeftPad = arg.input_left_pads_[i];
const index_t RightPad = arg.input_right_pads_[i];
if(!(X == 1 && LeftPad == 0 && RightPad == 0))
{
return false;
}
}
}
// check vector access of A
// FIXME: layout
if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
is_same_v<ALayout, ctc::NDHWGC>)
{
const index_t C = arg.a_g_n_c_wis_lengths_[2];
if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
{
return false;
}
}
else
{
return false;
}
// check vector access of B
// FIXME: layout
if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
is_same_v<BLayout, ctc::KZYXGC>)
{
const index_t C = arg.b_g_k_c_xs_lengths_[2];
if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
{
return false;
}
}
else
{
return false;
}
// check vector access of Ds
bool valid = true;
static_for<0, NumDTensor, 1>{}([&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
// FIXME: layout
if constexpr(is_same_v<DLayout, ctc::G_NW_K> || is_same_v<DLayout, ctc::G_NHW_K> ||
is_same_v<DLayout, ctc::G_NDHW_K> || is_same_v<DLayout, ctc::GNWK> ||
is_same_v<DLayout, ctc::GNHWK> || is_same_v<DLayout, ctc::GNDHWK> ||
is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::GK> ||
is_same_v<DLayout, ctc::G_K>)
{
const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
if(!(K % CDEShuffleBlockTransferScalarPerVector_NPerBlock == 0))
{
valid = false;
}
}
else
{
valid = false;
}
});
if(!valid)
{
return false;
}
// check vector access of E
if constexpr(is_same_v<ELayout, ctc::G_NW_K> || is_same_v<ELayout, ctc::G_NHW_K> ||
is_same_v<ELayout, ctc::G_NDHW_K> || is_same_v<ELayout, ctc::GNWK> ||
is_same_v<ELayout, ctc::GNHWK> || is_same_v<ELayout, ctc::GNDHWK> ||
is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
is_same_v<ELayout, ctc::NDHWGK>)
{
const index_t K = arg.e_g_n_k_wos_lengths_[2];
if(!(K % CDEShuffleBlockTransferScalarPerVector_NPerBlock == 0))
{
return false;
}
}
else
{
return false;
}
// check Gridwise GEMM
return GridwiseOp::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.ds_grid_desc_m_n_,
arg.e_grid_desc_m_n_,
arg.block_2_etile_map_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(
const void* p_a,
const void* p_b,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
const std::array<index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op)
{
return Argument{p_a,
p_b,
p_ds,
p_e,
a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
ds_g_n_k_wos_lengths,
ds_g_n_k_wos_strides,
e_g_n_k_wos_lengths,
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
a_element_op,
b_element_op,
cde_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseArgument> MakeArgumentPointer(
const void* p_a,
const void* p_b,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
const std::array<index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) override
{
return std::make_unique<Argument>(p_a,
p_b,
p_ds,
p_e,
a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
ds_g_n_k_wos_lengths,
ds_g_n_k_wos_strides,
e_g_n_k_wos_lengths,
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
a_element_op,
b_element_op,
cde_element_op);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceGroupedConvFwdMultipleD_Wmma_CShuffle"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< KPerBlock << ", "
<< getConvForwardSpecializationString(ConvForwardSpecialization)
<< ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
View file @ f8a6c69c
......@@ -10,46 +10,11 @@
#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_normalization_selector.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
template <typename GridwiseReduction,
typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
typename AccDataType,
typename AccElementwiseOperation,
typename GridDesc_M_K>
__global__ void kernel_normalization(const GridDesc_M_K x_grid_desc_m_k,
const GridDesc_M_K gamma_grid_desc_m_k,
const GridDesc_M_K beta_grid_desc_m_k,
const GridDesc_M_K y_grid_desc_m_k,
index_t num_k_block_tile_iteration,
AccDataType epsilon,
const XDataType* const __restrict__ p_x_global,
const GammaDataType* const __restrict__ p_gamma_global,
const BetaDataType* const __restrict__ p_beta_global,
YDataType* const __restrict__ p_y_global,
const AccElementwiseOperation acc_elementwise_op)
{
GridwiseReduction::Run(x_grid_desc_m_k,
gamma_grid_desc_m_k,
beta_grid_desc_m_k,
y_grid_desc_m_k,
num_k_block_tile_iteration,
epsilon,
p_x_global,
p_gamma_global,
p_beta_global,
p_y_global,
acc_elementwise_op);
};
} // namespace ck
namespace ck {
namespace tensor_operation {
namespace device {
......@@ -58,9 +23,9 @@ namespace device {
template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename AccDataType,
typename ComputeDataType,
typename YDataType,
typename AccElementwiseOperation,
typename YElementwiseOperation,
index_t Rank,
index_t NumReduceDim,
index_t BlockSize,
......@@ -74,16 +39,18 @@ template <typename XDataType,
index_t GammaSrcVectorSize,
index_t BetaSrcVectorDim,
index_t BetaSrcVectorSize,
index_t YDstVectorSize>
index_t YDstVectorSize,
bool UseWelford = true>
struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
ComputeDataType,
YDataType,
AccElementwiseOperation,
YElementwiseOperation,
Rank,
NumReduceDim>
{
static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize);
static_assert(
((GammaSrcVectorDim == 0 && MThreadSliceSize % GammaSrcVectorSize == 0) ||
(GammaSrcVectorDim == 1 && KThreadSliceSize % GammaSrcVectorSize == 0)),
......@@ -167,51 +134,6 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
using GridwiseReduceLayernormGeneric =
GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
AccElementwiseOperation,
GridDesc_M_K,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
XYSrcVectorDim,
XSrcVectorSize,
GammaSrcVectorDim,
GammaSrcVectorSize,
BetaSrcVectorDim,
BetaSrcVectorSize,
XYSrcVectorDim,
YDstVectorSize,
false>;
using GridwiseNormalizationSweepOnce =
GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
AccElementwiseOperation,
GridDesc_M_K,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
XYSrcVectorDim,
XSrcVectorSize,
GammaSrcVectorDim,
GammaSrcVectorSize,
BetaSrcVectorDim,
BetaSrcVectorSize,
XYSrcVectorDim,
YDstVectorSize,
true>;
struct Argument : public BaseArgument
{
Argument(const std::vector<index_t> lengths,
......@@ -220,7 +142,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
const std::vector<index_t> betaStrides,
const std::vector<index_t> yStrides,
const std::vector<index_t> reduceDims,
AccElementwiseOperation acc_elementwise_op,
YElementwiseOperation y_elementwise_op,
double epsilon,
const XDataType* p_x,
const GammaDataType* p_gamma,
......@@ -230,9 +152,9 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
p_gamma_(p_gamma),
p_beta_(p_beta),
p_y_(p_y),
acc_elementwise_op_(acc_elementwise_op)
y_elementwise_op_(y_elementwise_op)
{
epsilon_ = static_cast<AccDataType>(epsilon);
epsilon_ = static_cast<ComputeDataType>(epsilon);
Lengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(lengths, reduceDims);
xStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(xStrides, reduceDims);
......@@ -265,7 +187,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
}
AccDataType epsilon_;
ComputeDataType epsilon_;
const XDataType* p_x_;
const GammaDataType* p_gamma_;
......@@ -278,7 +200,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
std::vector<index_t> betaStrides_;
std::vector<index_t> yStrides_;
AccElementwiseOperation acc_elementwise_op_;
YElementwiseOperation y_elementwise_op_;
int blkGroupSize_;
int numBlockTileIteration_;
......@@ -295,23 +217,27 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
const auto kernel_main = arg.isSweeponce_
? kernel_normalization<GridwiseNormalizationSweepOnce,
XDataType,
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
AccElementwiseOperation,
GridDesc_M_K>
: kernel_normalization<GridwiseReduceLayernormGeneric,
XDataType,
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
AccElementwiseOperation,
GridDesc_M_K>;
auto kernel_main = NormalizationKernelSelector<XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
YElementwiseOperation,
GridDesc_M_K,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
XYSrcVectorDim,
XSrcVectorSize,
GammaSrcVectorDim,
GammaSrcVectorSize,
BetaSrcVectorDim,
BetaSrcVectorSize,
XYSrcVectorDim,
YDstVectorSize,
UseWelford>(arg.isSweeponce_);
float avg_time = 0;
avg_time += launch_and_time_kernel(stream_config,
......@@ -329,7 +255,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
arg.p_gamma_,
arg.p_beta_,
arg.p_y_,
arg.acc_elementwise_op_);
arg.y_elementwise_op_);
return (avg_time);
};
......@@ -429,7 +355,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
void* p_y,
void* p_saveMean,
void* p_saveInvVar,
AccElementwiseOperation acc_elementwise_op) override
YElementwiseOperation y_elementwise_op) override
{
// TODO
// Optional cache of the intermediate results (mean and InvVariance) during the
......@@ -443,7 +369,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
betaStrides,
yStrides,
reduceDims,
acc_elementwise_op,
y_elementwise_op,
epsilon,
static_cast<const XDataType*>(p_x),
static_cast<const GammaDataType*>(p_gamma),
......@@ -462,8 +388,8 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
// clang-format off
str << "DeviceNormalizationImpl<" << BlockSize << ",";
str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
str << "Cluster_MK_" << MThreadClusterSize << "_" << KThreadClusterSize << ",";
str << "Slice_MK_" << MThreadSliceSize << "_" << KThreadSliceSize << ",";
str << "XYSrcVectorDim_" << XYSrcVectorDim << ",";
str << "VectorSize_X" << XSrcVectorSize << "_Gamma" << GammaSrcVectorSize << "_Beta" << BetaSrcVectorSize << "_Y" << YDstVectorSize << ">";
// clang-format on
......
include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
View file @ f8a6c69c
......@@ -49,6 +49,14 @@ struct Add
y = x0 + x1;
};
template <>
__host__ __device__ constexpr void
operator()<float>(float& y, const float& x0, const bhalf_t& x1) const
{
const float x1_tmp = ck::type_convert<float>(x1);
y = x0 + x1_tmp;
}
template <>
__host__ __device__ constexpr void
operator()<bhalf_t>(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const
......@@ -67,6 +75,30 @@ struct Add
};
};
struct ScaleAdd
{
__host__ __device__ ScaleAdd(float scale) : scale_(scale) {}
template <typename Y, typename X0, typename X1>
__host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
template <>
__host__ __device__ void
operator()<float, float, half_t>(float& y, const float& x0, const half_t& x1) const
{
y = scale_ * x0 + ck::type_convert<float>(x1);
};
template <>
__host__ __device__ void
operator()<float, float, bhalf_t>(float& y, const float& x0, const bhalf_t& x1) const
{
y = scale_ * x0 + ck::type_convert<float>(x1);
};
float scale_;
};
struct Subtract
{
template <typename T>
......@@ -118,6 +150,13 @@ struct Bilinear
template <typename Y, typename X0, typename X1>
__host__ __device__ constexpr void operator()(Y&, const X0&, const X1&) const;
template <>
__host__ __device__ constexpr void
operator()<double, double, double>(double& y, const double& x0, const double& x1) const
{
y = alpha_ * x0 + beta_ * x1;
};
template <>
__host__ __device__ constexpr void
operator()<float, float, float>(float& y, const float& x0, const float& x1) const
......
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ f8a6c69c
......@@ -95,6 +95,12 @@ struct Scale
y = scale_ * x;
};
template <>
__host__ __device__ void operator()<double, double>(double& y, const double& x) const
{
y = scale_ * x;
};
float scale_;
};
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp 0 → 100644
View file @ f8a6c69c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/multi_index_transform_helper.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_softmax.hpp"
namespace ck {
template <typename FloatAB,
typename FloatGemmAcc,
typename FloatCShuffle,
typename FloatC,
typename D0sDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename C0DEElementwiseOperation,
typename B1ElementwiseOperation,
typename C1DEElementwiseOperation,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename B1GridDesc_BK0_N_BK1,
typename C1GridDesc_M_N,
typename D0sGridDesc_M_N,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t Gemm1NPerBlock,
index_t Gemm1KPerBlock,
index_t AK1Value,
index_t BK1Value,
index_t B1K1Value,
index_t MPerXdl,
index_t NPerXdl,
index_t MXdlPerWave,
index_t NXdlPerWave,
index_t Gemm1NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool AThreadTransferSrcResetCoordinateAfterRun, // ignored
index_t ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
index_t BBlockLdsExtraN,
typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
typename B1BlockTransferThreadClusterArrangeOrder,
typename B1BlockTransferSrcAccessOrder,
index_t B1BlockTransferSrcVectorDim,
index_t B1BlockTransferSrcScalarPerVector,
index_t B1BlockTransferDstScalarPerVector_BK1,
bool B1ThreadTransferSrcResetCoordinateAfterRun,
index_t B1BlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
LoopScheduler LoopSched,
bool PadN,
bool MaskOutUpperTriangle,
PipelineVersion PipelineVer = PipelineVersion::v1>
struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle
{
static_assert(LoopSched == LoopScheduler::Default,
"Non-default loop scheduler is currently not supported");
static constexpr index_t NumD0Tensor = D0sDataType::Size();
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>{};
static constexpr auto I5 = Number<5>{};
static constexpr auto I6 = Number<6>{};
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
// Gemm0
static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
static constexpr auto AK1 = Number<AK1Value>{};
static constexpr auto BK1 = Number<BK1Value>{};
static constexpr auto Gemm0MWaves = MPerBlock / (MPerXdl * MXdlPerWave);
static constexpr auto Gemm0NWaves = NPerBlock / (NPerXdl * NXdlPerWave);
// Gemm1
static constexpr auto B1K0 = Number<Gemm1KPerBlock / B1K1Value>{};
static constexpr auto B1K1 = Number<B1K1Value>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
using GridwiseGemmPipe = remove_cvref_t<decltype(
GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
template <typename ABlockDesc_AK0_M_AK1>
__host__ __device__ static constexpr auto
MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
{
constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, MWaves, MPerXdl>(
ABlockDesc_AK0_M_AK1{});
}
template <typename BBlockDesc_BK0_N_BK1>
__host__ __device__ static constexpr auto
MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
{
constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
BBlockDesc_BK0_N_BK1{});
}
template <typename ABlockDesc_AK0_M_AK1>
__host__ __device__ static constexpr auto
MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
{
return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, 1, 1>(ABlockDesc_AK0_M_AK1{});
}
template <typename BBlockDesc_BK0_N_BK1>
__host__ __device__ static constexpr auto
MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
{
constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, NPerXdl>(
BBlockDesc_BK0_N_BK1{});
}
__host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
{
// A matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(AK0, Number<MPerBlock>{}, AK1),
make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
}
__host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
{
// B matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(BK0, Number<NPerBlock>{}, BK1),
make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
}
__host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
{
// B1 matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(B1K0, Number<Gemm1NPerBlock>{}, B1K1),
make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
}
__host__ __device__ static constexpr auto
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
{
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
I1,
Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
}
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
const index_t gemm0_bytes_end = (SharedMemTrait::a_block_space_size_aligned +
SharedMemTrait::b_block_space_size_aligned) *
sizeof(FloatAB);
const index_t gemm1_bytes_end =
(SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
sizeof(FloatAB);
const index_t softmax_bytes_end = (SharedMemTrait::reduction_space_offset +
SharedMemTrait::reduction_space_size_aligned) *
sizeof(FloatGemmAcc);
const index_t c_block_bytes_end =
SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
return math::max(gemm0_bytes_end, gemm1_bytes_end, softmax_bytes_end, c_block_bytes_end);
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
template <typename Block2CTileMap>
__host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
const C1GridDesc_M_N& c1_grid_desc_m_n,
const Block2CTileMap& block_2_ctile_map)
{
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
"Invalid tuning param!");
const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
const auto Gemm1N = b1_grid_desc_bk0_n_bk1.GetLength(I1);
if(!(M == c1_grid_desc_m_n.GetLength(I0) && Gemm1N == c1_grid_desc_m_n.GetLength(I1)))
{
return false;
}
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0 &&
Gemm1N % Gemm1NPerBlock == 0))
{
return false;
}
// check gemm0 gridwise gemm pipeline
const auto num_gemm0_k_loop = K / KPerBlock;
if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
{
return false;
}
// check gemm1 gridwise gemm pipeline
if(!(NPerBlock % Gemm1KPerBlock == 0))
{
return false;
}
const auto num_gemm1_k_inner_loop = NPerBlock / Gemm1KPerBlock;
if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
{
return false;
}
if(!block_2_ctile_map.CheckValidity(c1_grid_desc_m_n))
{
return false;
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return true;
}
__host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
{
const index_t num_loop = K / KPerBlock;
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
}
__host__ __device__ static constexpr auto
MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const C1GridDesc_M_N& c1_grid_desc_m_n)
{
const auto M = c1_grid_desc_m_n.GetLength(I0);
const auto N = c1_grid_desc_m_n.GetLength(I1);
const auto MBlock = M / MPerBlock;
const auto NBlock = N / Gemm1NPerBlock;
const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
c1_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
make_unmerge_transform(make_tuple(NBlock, Number<Gemm1NPerBlock>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
return c_grid_desc_mblock_mperblock_nblock_nperblock;
}
// return block_id to C matrix tile idx (m0, n0) mapping
__host__ __device__ static constexpr auto
MakeDefaultBlock2CTileMap(const C1GridDesc_M_N& c1_grid_desc_m_n)
{
return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, C1GridDesc_M_N>(
c1_grid_desc_m_n);
}
__device__ static auto GetGemm0WaveIdx()
{
const index_t thread_id = get_thread_local_1d_id();
constexpr auto WaveSize = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.wave_size;
constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(Gemm0MWaves, Gemm0NWaves, WaveSize))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
}
__device__ static auto GetGemm0WaveMNIdx(const index_t thread_id)
{
constexpr auto WaveSize = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.wave_size;
constexpr auto wave_threadid_to_mn_idx_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(WaveSize / MPerXdl, MPerXdl))),
make_tuple(Sequence<0, 1>{}),
make_tuple(Sequence<0>{}));
return wave_threadid_to_mn_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
}
static constexpr auto MakeD0sGridPointer()
{
return generate_tuple(
[&](auto i) {
using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
return static_cast<const D0DataType*>(nullptr);
},
Number<NumD0Tensor>{});
}
// D0 desc for source in blockwise copy
template <typename D0GridDesc_M_N>
__host__ __device__ static constexpr auto
MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0GridDesc_M_N& d0_grid_desc_m_n)
{
const auto M = d0_grid_desc_m_n.GetLength(I0);
const auto N = d0_grid_desc_m_n.GetLength(I1);
constexpr auto mfma = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma;
constexpr auto N3 = mfma.num_groups_per_blk;
constexpr auto N4 = mfma.num_input_blks;
constexpr auto N5 = mfma.group_size;
return transform_tensor_descriptor(
d0_grid_desc_m_n,
make_tuple(make_unmerge_transform(
make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)),
make_unmerge_transform(
make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
}
// D0s desc for source in blockwise copy
__host__ __device__ static constexpr auto
MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0sGridDesc_M_N& ds_grid_desc_m_n)
{
return generate_tuple(
[&](auto i) {
return MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ds_grid_desc_m_n[i]);
},
Number<NumD0Tensor>{});
}
using D0sGridPointer = decltype(MakeD0sGridPointer());
using D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(D0sGridDesc_M_N{}))>;
using C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(C1GridDesc_M_N{}))>;
using DefaultBlock2CTileMap =
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(C1GridDesc_M_N{}))>;
struct SharedMemTrait
{
// LDS allocation for A and B: be careful of alignment
static constexpr auto a_block_desc_ak0_m_ak1 =
GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
static constexpr auto b_block_desc_bk0_n_bk1 =
GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
static constexpr auto b1_block_desc_bk0_n_bk1 =
GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
static constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
static constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
static constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
static constexpr auto a_block_space_offset = 0;
static constexpr auto b_block_space_offset = a_block_space_size_aligned.value;
static constexpr auto b1_block_space_offset = 0;
// LDS allocation for reduction
static constexpr index_t reduction_space_size_aligned =
math::integer_least_multiple(BlockSize, max_lds_align);
static constexpr auto reduction_space_offset = 0;
// LDS allocation for C shuffle in LDS
static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
static constexpr auto c_block_space_size =
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
};
template <bool HasMainKBlockLoop, typename Block2CTileMap, typename C0MatrixMask>
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
const FloatAB* __restrict__ p_b1_grid,
FloatC* __restrict__ p_c_grid,
D0sGridPointer p_d0s_grid,
void* __restrict__ p_shared,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const C0DEElementwiseOperation& c0de_element_op,
const B1ElementwiseOperation& b1_element_op,
const C1DEElementwiseOperation& c1de_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 B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock,
const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
const Block2CTileMap& block_2_ctile_map,
const C0MatrixMask& c0_matrix_mask)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
const auto d0s_grid_buf = generate_tuple(
[&](auto i) {
return make_dynamic_buffer<AddressSpaceEnum::Global>(
p_d0s_grid[i],
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i].GetElementSpaceSize());
},
Number<NumD0Tensor>{});
// divide block work by [M, N]
const auto block_work_idx =
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_ctile_map.ValidCTileIndex(
block_work_idx,
make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
// HACK: this force m/gemm1_n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
const index_t gemm1_n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
//
// set up Gemm0
//
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<AK0, MPerBlock, AK1>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(a_grid_desc_ak0_m_ak1),
decltype(a_block_desc_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
Sequence<1, 0, 2>,
ABlockTransferSrcVectorDim,
2,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
1,
1,
true, // SrcResetCoord
true, // DstResetCoord
NumGemmKPrefetchStage>(
a_grid_desc_ak0_m_ak1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_ak0_m_ak1,
make_multi_index(0, 0, 0),
tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<BK0, NPerBlock, BK1>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_block_desc_bk0_n_bk1),
BBlockTransferSrcAccessOrder,
Sequence<1, 0, 2>,
BBlockTransferSrcVectorDim,
2,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
1,
1,
true, // SrcResetCoord
true, // DstResetCoord
NumGemmKPrefetchStage>(
b_grid_desc_bk0_n_bk1,
make_multi_index(0, 0, 0), // will loop over GemmN dimension
b_element_op,
b_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0),
tensor_operation::element_wise::PassThrough{});
// Fused Gemm+Gemm pipeline
// for n in N0:
// for k in K0:
// acc[m][n] += A[m][k] * B0[k][n]
// acc1[m][o] += acc[m][n] * B1[n][o]
// sanity check
constexpr index_t KPack = math::max(
math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
auto blockwise_gemm = BlockwiseGemmXdlops_v2<
BlockSize,
FloatAB,
FloatGemmAcc,
decltype(a_block_desc_ak0_m_ak1),
decltype(b_block_desc_bk0_n_bk1),
decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
MPerBlock,
NPerBlock,
KPerBlock,
MPerXdl,
NPerXdl,
MXdlPerWave,
NXdlPerWave,
KPack,
true>{}; // TransposeC
auto acc_thread_buf = blockwise_gemm.GetCThreadBuffer();
// LDS allocation for A and B: be careful of alignment
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + SharedMemTrait::a_block_space_offset,
a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + SharedMemTrait::b_block_space_offset,
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
const auto a_block_reset_copy_step =
make_multi_index(-a_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
const auto b_block_reset_copy_step =
make_multi_index(-b_grid_desc_bk0_n_bk1.GetLength(I0), NPerBlock, 0);
// gridwise GEMM pipeline
// Only supports LoopScheduler::Default
const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_Selector<PipelineVer,
NumGemmKPrefetchStage,
LoopScheduler::Default>();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
//
// set up Gemm1
//
// Acc matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
constexpr auto acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
constexpr auto m0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
constexpr auto n0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
constexpr auto m1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
constexpr auto n1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
constexpr auto m2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
constexpr auto n2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
constexpr auto n3 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
constexpr auto n4 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
// d0 matrix threadwise copy
constexpr auto d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
make_naive_tensor_descriptor_packed(make_tuple(I1, // MBlockId
I1, // NBlockID
I1, // MRepeat
I1, // NRepeat
I1, // MWaveId
I1, // NWaveId
I1, // MPerXdl
I1, // NGroupNum
I1, // NInputNum
n4)); // registerNum
auto d0s_thread_buf = generate_tuple(
[&](auto i) {
using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
return StaticBuffer<
AddressSpaceEnum::Vgpr,
D0DataType,
d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(),
true>{};
},
Number<NumD0Tensor>{});
const auto wave_id = GetGemm0WaveIdx();
const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
constexpr auto acc0_thread_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<MXdlPerWave>{}, Number<NXdlPerWave>{}, n2, n4));
auto d0s_threadwise_copy = generate_tuple(
[&](auto i) {
using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
return ThreadwiseTensorSliceTransfer_v2<
D0DataType,
D0DataType,
decltype(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i]),
decltype(d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
Sequence<I1, I1, I1, I1, I1, I1, I1, I1, I1, n4>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
9,
n4,
1,
false>(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
make_multi_index(block_work_idx[I0], // MBlockId
0, // NBlockId
0, // mrepeat
0, // nrepeat
wave_id[I0], // MWaveId
wave_id[I1], // NWaveId
wave_m_n_id[I1], // MPerXdl
0, // group
wave_m_n_id[I0], // NInputIndex
0)); // register number
},
Number<NumD0Tensor>{});
// acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc_thread_desc_k0_m_k1
// n0_n1_n2_n3 -> k0
// m0_m1_m2 -> m
// n4 -> k1
// NOTE: had to use merge_v3 or will spit out compilation errors
constexpr auto acc_thread_desc_k0_m_k1 = transform_tensor_descriptor(
acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
make_pass_through_transform(n4)),
make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
// A1 matrix in AccVGPR
// N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
constexpr auto AccN3 =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
constexpr auto A1ThreadSlice_K0_M_K1 =
make_tuple(Number<Gemm1KPerBlock / n4 / AccN3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
constexpr auto A1ThreadSliceK0 = A1ThreadSlice_K0_M_K1[I0];
constexpr auto A1ThreadSliceM = A1ThreadSlice_K0_M_K1[I1];
constexpr auto A1ThreadSliceK1 = A1ThreadSlice_K0_M_K1[I2];
constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
A1ThreadSlice_K0_M_K1,
make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
// B1 matrix in LDS memory, dst of blockwise copy
constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// A1 matrix blockwise copy
auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
FloatGemmAcc,
FloatAB,
decltype(acc_thread_desc_k0_m_k1),
decltype(a1_thread_desc_k0_m_k1),
tensor_operation::element_wise::PassThrough,
Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
Sequence<1, 0, 2>,
2,
n4>{tensor_operation::element_wise::PassThrough{}};
// B1 matrix blockwise copy
auto b1_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<B1K0, Gemm1NPerBlock, B1K1>,
B1BlockTransferThreadClusterLengths_BK0_N_BK1,
B1BlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(b1_grid_desc_bk0_n_bk1),
decltype(b1_block_desc_bk0_n_bk1),
B1BlockTransferSrcAccessOrder,
Sequence<1, 0, 2>,
B1BlockTransferSrcVectorDim,
2,
B1BlockTransferSrcScalarPerVector,
B1BlockTransferDstScalarPerVector_BK1,
1,
1,
B1ThreadTransferSrcResetCoordinateAfterRun,
true, // DstResetCoord
NumGemmKPrefetchStage>(
b1_grid_desc_bk0_n_bk1,
make_multi_index(0, gemm1_n_block_data_idx_on_grid, 0),
b1_element_op,
b1_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0),
tensor_operation::element_wise::PassThrough{});
auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
a1_thread_desc_k0_m_k1.GetElementSpaceSize());
// reuse LDS space for gemm0's b_block_buf
auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + SharedMemTrait::b1_block_space_offset,
b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
// selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
// selected_mfma.k_per_blk <= Gemm1KPack
//
// Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
// multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
// Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
// with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
// cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
// therefore we may just as well assign Gemm1KPack = group_size
constexpr index_t Gemm1KPack =
MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
auto gemm1_blockwise_gemm = BlockwiseGemmXdlops_v2<
BlockSize,
FloatAB,
FloatGemmAcc,
decltype(a1_thread_desc_k0_m_k1),
decltype(b1_block_desc_bk0_n_bk1),
decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
MPerBlock,
Gemm1NPerBlock,
Gemm1KPerBlock,
MPerXdl,
NPerXdl,
MXdlPerWave,
Gemm1NXdlPerWave,
Gemm1KPack,
true, // TransposeC
Gemm1KPack, // AMmaKStride
Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
// BMmaKStride
make_tuple(0, 0, 0, 0)}; // A_origin
auto acc1_thread_buf = gemm1_blockwise_gemm.GetCThreadBuffer();
//
// Blockwise softmax
//
auto workspace_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatGemmAcc*>(p_shared) + SharedMemTrait::reduction_space_offset,
SharedMemTrait::reduction_space_size_aligned);
// get acc0 8D thread cluster
constexpr auto thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4 =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths() /
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
constexpr auto tm0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I0);
constexpr auto tn0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I1);
constexpr auto tm1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I2);
constexpr auto tn1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I3);
constexpr auto tm2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I4);
constexpr auto tn2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I5);
constexpr auto tn3 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I6);
constexpr auto tn4 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I7);
// get acc0 thread map
constexpr auto m0_n_m1_to_m_n_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(tm0 * tm1, tm2)),
make_pass_through_transform(I1)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
constexpr auto threadid_to_m0_n_m1_adaptor = make_single_stage_tensor_adaptor(
make_tuple(
make_merge_transform(make_tuple(tm0 * tm1, tn0 * tn1 * tn2 * tn3 * tn4, tm2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto threadid_to_m_n_thread_cluster_adaptor =
chain_tensor_adaptors(m0_n_m1_to_m_n_adaptor, threadid_to_m0_n_m1_adaptor);
// get acc0 2D thread cluster & 2D thread slice
constexpr auto thread_cluster_desc_m_n = make_naive_tensor_descriptor_packed(
make_tuple(tm0 * tm1 * tm2, tn0 * tn1 * tn2 * tn3 * tn4));
constexpr auto thread_slice_desc_m_n =
make_naive_tensor_descriptor_packed(make_tuple(m0 * m1 * m2, n0 * n1 * n2 * n3 * n4));
auto blockwise_softmax = BlockwiseSoftmax<BlockSize,
FloatGemmAcc,
decltype(threadid_to_m_n_thread_cluster_adaptor),
decltype(thread_cluster_desc_m_n),
decltype(thread_slice_desc_m_n)>{};
const index_t num_gemm1_k_block_outer_loop =
b_grid_desc_bk0_n_bk1.GetLength(I1) / NPerBlock;
constexpr index_t num_gemm1_k_block_inner_loop = NPerBlock / Gemm1KPerBlock;
// Initialize C
StaticBuffer<AddressSpaceEnum::Vgpr, FloatGemmAcc, acc1_thread_buf.Size(), true>
c_thread_buf;
c_thread_buf.Clear();
// Initialize running sum and max of exponentiating row vectors
using SoftmaxBuf = typename decltype(blockwise_softmax)::BufferType;
SoftmaxBuf running_sum, running_sum_new, running_max, running_max_new;
running_sum = 0;
running_sum_new = 0;
running_max = NumericLimits<FloatGemmAcc>::Lowest();
running_max_new = NumericLimits<FloatGemmAcc>::Lowest();
// gemm1 K loop
index_t gemm1_k_block_outer_index = 0;
do
{
auto n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(gemm1_k_block_outer_index * NPerBlock);
if(c0_matrix_mask.IsTileSkippable(
m_block_data_idx_on_grid, n_block_data_idx_on_grid, MPerBlock, NPerBlock))
{
continue;
}
// gemm0
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
a_block_desc_ak0_m_ak1,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_grid_desc_bk0_n_bk1,
b_block_desc_bk0_n_bk1,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
blockwise_gemm,
acc_thread_buf,
num_k_block_main_loop);
// multiple d
if constexpr(NumD0Tensor)
{
static_for<0, MXdlPerWave, 1>{}([&](auto mr) {
static_for<0, NXdlPerWave, 1>{}([&](auto nr) {
static_for<0, n2, 1>{}([&](auto groupid) {
static_for<0, NumD0Tensor, 1>{}([&](auto i) {
d0s_threadwise_copy(i).Run(
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
d0s_grid_buf[i],
d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
d0s_thread_buf(i));
});
static_for<0, n4, 1>{}([&](auto i) {
constexpr index_t c_offset = acc0_thread_desc.CalculateOffset(
make_tuple(mr, nr, groupid, i));
// get reference to src data
const auto src_data_refs = generate_tie(
// return type should be lvalue
[&](auto iSrc) -> const auto& {
return d0s_thread_buf[iSrc][i];
},
Number<NumD0Tensor>{});
// get reference to dst data
auto dst_data_refs = generate_tie(
// return type should be lvalue
[&](auto) -> auto& {
return acc_thread_buf(Number<c_offset>{});
},
Number<2>{});
unpack2(c0de_element_op, dst_data_refs, src_data_refs);
});
static_for<0, NumD0Tensor, 1>{}([&](auto i) {
d0s_threadwise_copy(i).MoveSrcSliceWindow(
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
make_multi_index(0, 0, 0, 0, 0, 0, 0, 1, 0, 0));
});
});
static_for<0, NumD0Tensor, 1>{}([&](auto i) {
d0s_threadwise_copy(i).MoveSrcSliceWindow(
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
make_multi_index(0, 0, 0, 1, 0, 0, 0, -n2.value, 0, 0));
});
});
static_for<0, NumD0Tensor, 1>{}([&](auto i) {
d0s_threadwise_copy(i).MoveSrcSliceWindow(
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
make_multi_index(0, 0, 1, -NXdlPerWave, 0, 0, 0, 0, 0, 0));
});
});
static_for<0, NumD0Tensor, 1>{}([&](auto i) {
d0s_threadwise_copy(i).MoveSrcSliceWindow(
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
make_multi_index(0, 1, -MXdlPerWave, 0, 0, 0, 0, 0, 0, 0));
});
}
else
{
static_for<0, acc_thread_buf.Size(), 1>{}(
[&](auto i) { c0de_element_op(acc_thread_buf(i), acc_thread_buf[i]); });
}
// do MNK padding or upper triangular masking
if constexpr(MaskOutUpperTriangle || PadN)
{
// 8d thread_desc in thread scope
constexpr auto c_thread_lengths =
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
// 8d block_desc in block scope
constexpr auto c_block_lengths =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
constexpr auto M0 = c_block_lengths[I0];
constexpr auto N0 = c_block_lengths[I1];
constexpr auto M1 = c_block_lengths[I2];
constexpr auto N1 = c_block_lengths[I3];
constexpr auto M2 = c_block_lengths[I4];
constexpr auto N2 = c_block_lengths[I5];
constexpr auto N3 = c_block_lengths[I6];
constexpr auto N4 = c_block_lengths[I7];
// works like multi-dimension static_for (static_ford), but provides both the linear
// index as well as n-d index
using Acc0TileIterator = SpaceFillingCurve<
decltype(c_thread_lengths),
typename arithmetic_sequence_gen<0, c_thread_lengths.Size(), 1>::type,
typename uniform_sequence_gen<c_thread_lengths.Size(), 1>::type,
false>; // SnakeCurved
auto acc0_thread_origin = blockwise_gemm.CalculateCThreadOriginDataIndex8D(
Number<0>{}, Number<0>{}, Number<0>{}, Number<0>{});
constexpr auto block_idx_to_m_n_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(M0, M1, M2)),
make_unmerge_transform(make_tuple(N0, N1, N2, N3, N4))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5, 6, 7>{}));
static_for<0, Acc0TileIterator::GetNumOfAccess(), 1>{}([&](auto i) {
auto acc0_thread_idx = Acc0TileIterator::GetIndex(i) + acc0_thread_origin;
auto m_local =
block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I0];
auto n_local =
block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I1];
auto m_global = m_local + m_block_data_idx_on_grid;
auto n_global = n_local + n_block_data_idx_on_grid;
if(c0_matrix_mask.IsMaskedElement(m_global, n_global))
{
acc_thread_buf(i) = -ck::NumericLimits<float>::Infinity();
}
});
}
block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
// softmax
SoftmaxBuf& max = blockwise_softmax.max_value_buf;
SoftmaxBuf& sum = blockwise_softmax.sum_value_buf;
blockwise_softmax.Run(acc_thread_buf, workspace_buf);
// TODO: may convert to log domain
running_max_new = mathext::max(max, running_max);
running_sum_new = mathext::exp(running_max - running_max_new) * running_sum +
mathext::exp(max - running_max_new) * sum;
// gemm1
{
// TODO: explore using dynamic buffer for a1 thread buffer
// For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
// RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
// the A1 source buffer is static buffer holding the output of first GEMM and
// requires constexpr offset by design. Therefore, we pass tensor coordinate offset
// explicitly in Run() below.
// Initialize acc1
acc1_thread_buf.Clear();
// preload data into LDS
b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
b1_block_slice_copy_step);
block_sync_lds(); // wait for reduction LDS read
b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
// main body
if constexpr(num_gemm1_k_block_inner_loop > 1)
{
static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
a1_blockwise_copy.Run(acc_thread_desc_k0_m_k1,
make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
acc_thread_buf,
a1_thread_desc_k0_m_k1,
make_tuple(I0, I0, I0),
a1_thread_buf);
b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
block_sync_lds();
gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
block_sync_lds();
b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
b1_block_slice_copy_step);
b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
});
}
// tail
{
a1_blockwise_copy.Run(
acc_thread_desc_k0_m_k1,
make_tuple(
Number<(num_gemm1_k_block_inner_loop - 1) * A1ThreadSliceK0>{}, I0, I0),
acc_thread_buf,
a1_thread_desc_k0_m_k1,
make_tuple(I0, I0, I0),
a1_thread_buf);
block_sync_lds();
gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
}
} // end gemm1
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
constexpr auto cn0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
constexpr auto cm1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
constexpr auto cn1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
constexpr auto cm2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
constexpr auto cn2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
constexpr auto cn3 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
constexpr auto cn4 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
constexpr auto c_thread_slice_desc_m_n = make_naive_tensor_descriptor_packed(
make_tuple(cm0 * cm1 * cm2, cn0 * cn1 * cn2 * cn3 * cn4));
constexpr auto c_thread_buf_slice_m = c_thread_slice_desc_m_n.GetLength(I0);
constexpr auto c_thread_buf_slice_n = c_thread_slice_desc_m_n.GetLength(I1);
static_for<0, c_thread_buf_slice_m, 1>{}([&](auto iM) {
static_for<0, c_thread_buf_slice_n, 1>{}([&](auto iN) {
auto I = Number<c_thread_slice_desc_m_n.CalculateOffset(make_tuple(iM, iN))>{};
FloatGemmAcc acc1 = acc1_thread_buf[I]; // P*V
FloatGemmAcc c = c_thread_buf[I]; // O
FloatGemmAcc c_new =
(running_sum[iM] * math::exp(running_max[iM] - running_max_new[iM]) * c +
math::exp(max[iM] - running_max_new[iM]) * acc1) /
running_sum_new[iM]; // Formula by Dao et al.,
// https://arxiv.org/pdf/2205.14135v2.pdf section 3.1
c_thread_buf(I) = c_new; // O_new
});
});
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_ak0_m_ak1,
a_block_reset_copy_step); // rewind K
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_bk0_n_bk1,
b_block_reset_copy_step); // rewind K and step N
// update before next j iteration
running_max = running_max_new;
running_sum = running_sum_new;
block_sync_lds(); // wait for gemm1 LDS read
} while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
// shuffle C and write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
Gemm1NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp =
gemm1_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5);
constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6);
constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatCShuffle*>(p_shared),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2)), // M2 = MPerXdl
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2, // N2 * N3 * N4 = NPerXdl
N3,
N4))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
gemm1_blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
FloatCShuffle,
decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
decltype(c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
I1,
N2,
I1,
N4>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
n_thread_data_on_block_idx[I2],
n_thread_data_on_block_idx[I3],
n_thread_data_on_block_idx[I4]),
tensor_operation::element_wise::PassThrough{}};
// shuffle: blockwise copy C from LDS to global
auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
ThisThreadBlock, // ThreadGroup
C1DEElementwiseOperation, // ElementwiseOperation,
CGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
FloatCShuffle, // typename SrcData,
FloatC, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
true, // bool ThreadTransferSrcResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
c1de_element_op};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, 1, N2, 1, N4>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
1,
N2,
1,
N4>>{};
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_c_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, Gemm1NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
c_shuffle_block_copy_lds_to_global.Run(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
c_shuffle_block_buf,
c_grid_desc_mblock_mperblock_nblock_nperblock,
c_grid_buf);
if constexpr(access_id < num_access - 1)
{
constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
// move on C
c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
}
});
}
}
};
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
View file @ f8a6c69c
......@@ -879,14 +879,6 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
}
} // end gemm1
// workaround compiler issue; see ck/ck.hpp
if constexpr(CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE == 1 &&
is_same_v<FloatAB, bhalf_t> && MPerBlock == 256 && NPerBlock == 128 &&
Gemm1NPerBlock == 128)
{
__builtin_amdgcn_sched_barrier(0);
}
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp
View file @ f8a6c69c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -574,4 +574,546 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
}
};
template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_B_K0_M_K1,
typename BGridDesc_B_K0_N_K1,
typename CGridDesc_M_N,
index_t MPerBlock,
index_t NPerBlock,
index_t K0PerBlock,
index_t K1Value,
index_t M1PerThreadM111,
index_t N1PerThreadN111,
index_t KPerThread,
typename M11N11ThreadClusterM110Xs,
typename M11N11ThreadClusterN110Xs,
typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
typename ABlockTransferSrcVectorTensorContiguousDimOrder,
typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
typename BBlockTransferSrcVectorTensorContiguousDimOrder,
typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
typename CThreadTransferSrcDstAccessOrder,
index_t CThreadTransferSrcDstVectorDim,
index_t CThreadTransferDstScalarPerVector>
struct GridwiseGemmDl_bkm_bkn_mn_v1r3
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
// K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{};
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
// TODO: change this. I think it needs multi-dimensional alignment
constexpr auto max_lds_align = K1;
// TODO: check alignment
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_b_k0_m_k1 = make_naive_tensor_descriptor_aligned(
make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
// TODO: check alignment
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_b_k0_n_k1 = make_naive_tensor_descriptor_aligned(
make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
// TODO: check alignment
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
a_block_desc_b_k0_m_k1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
b_block_desc_b_k0_n_k1.GetElementSpaceSize(), max_lds_align);
return 2 * (a_block_aligned_space_size + b_block_aligned_space_size) * sizeof(FloatAB);
}
__host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_B_K0_M_K1& a_grid_desc_b_k0_m_k1,
const BGridDesc_B_K0_N_K1& b_grid_desc_b_k0_n_k1,
const CGridDesc_M_N& c_grid_desc_m_n)
{
const auto M = a_grid_desc_b_k0_m_k1.GetLength(I2);
const auto N = b_grid_desc_b_k0_n_k1.GetLength(I2);
const auto K0 = a_grid_desc_b_k0_m_k1.GetLength(I1);
const auto KBatch = a_grid_desc_b_k0_m_k1.GetLength(I0);
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return (M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
K0 == b_grid_desc_b_k0_n_k1.GetLength(I1) &&
K1 == a_grid_desc_b_k0_m_k1.GetLength(I3) &&
K1 == b_grid_desc_b_k0_n_k1.GetLength(I3)) &&
KBatch == b_grid_desc_b_k0_n_k1.GetLength(I0) &&
(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0);
}
__host__ __device__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
{
const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
return grid_size;
}
__host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K0)
{
const bool has_main_k_block_loop = (K0 + K0PerBlock) / (2 * K0PerBlock) > 1;
return has_main_k_block_loop;
}
__host__ __device__ static constexpr bool CalculateHasDoubleTailKBlockLoop(index_t K0)
{
const bool has_double_tail_k_block_loop = (K0 / K0PerBlock) % 2 == 0;
return has_double_tail_k_block_loop;
}
__host__ __device__ static constexpr auto
MakeAGridDescriptor_B_K0_M0_M1_K1(const AGridDesc_B_K0_M_K1& a_grid_desc_b_k0_m_k1)
{
const auto KBatch = a_grid_desc_b_k0_m_k1.GetLength(I0);
const auto K0 = a_grid_desc_b_k0_m_k1.GetLength(I1);
const auto M = a_grid_desc_b_k0_m_k1.GetLength(I2);
const auto M1 = Number<MPerBlock>{};
const auto M0 = M / M1;
const auto a_grid_desc_b_k0_m0_m1_k1 = transform_tensor_descriptor(
a_grid_desc_b_k0_m_k1,
make_tuple(make_pass_through_transform(KBatch),
make_pass_through_transform(K0),
make_unmerge_transform(make_tuple(M0, M1)),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}));
return a_grid_desc_b_k0_m0_m1_k1;
}
__host__ __device__ static constexpr auto
MakeBGridDescriptor_B_K0_N0_N1_K1(const BGridDesc_B_K0_N_K1& b_grid_desc_b_k0_n_k1)
{
const auto KBatch = b_grid_desc_b_k0_n_k1.GetLength(I0);
const auto K0 = b_grid_desc_b_k0_n_k1.GetLength(I1);
const auto N = b_grid_desc_b_k0_n_k1.GetLength(I2);
const auto N1 = Number<NPerBlock>{};
const auto N0 = N / N1;
const auto b_grid_desc_b_k0_n0_n1_k1 = transform_tensor_descriptor(
b_grid_desc_b_k0_n_k1,
make_tuple(make_pass_through_transform(KBatch),
make_pass_through_transform(K0),
make_unmerge_transform(make_tuple(N0, N1)),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}));
return b_grid_desc_b_k0_n0_n1_k1;
}
__host__ __device__ static constexpr auto
MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(const CGridDesc_M_N& c_grid_desc_m_n)
{
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;
constexpr auto M11 =
Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies{}, I1) *
M1PerThreadM111>{};
constexpr auto N11 =
Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies{}, I1) *
N1PerThreadN111>{};
constexpr auto M10 = M1 / M11;
constexpr auto N10 = N1 / N11;
const auto c_grid_desc_m0_m10_m11_n0_n10_n11 = transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(M0, M10, M11)),
make_unmerge_transform(make_tuple(N0, N10, N11))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
return c_grid_desc_m0_m10_m11_n0_n10_n11;
}
// return block_id to C matrix tile idx (m0, n0) mapping
__host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
const CGridDesc_M_N& c_m_n_grid_desc, index_t M01, index_t N01, index_t KBatch)
{
return BlockToCTileMap_KSplit_M00_N00_M01_N01<MPerBlock, NPerBlock, CGridDesc_M_N>(
c_m_n_grid_desc, M01, N01, KBatch);
}
using AGridDesc_B_K0_M0_M1_K1 =
decltype(MakeAGridDescriptor_B_K0_M0_M1_K1(AGridDesc_B_K0_M_K1{}));
using BGridDesc_B_K0_N0_N1_K1 =
decltype(MakeBGridDescriptor_B_K0_N0_N1_K1(BGridDesc_B_K0_N_K1{}));
using CGridDesc_M0_M10_M11_N0_N10_N11 =
decltype(MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
using CBlockClusterAdaptor = decltype(MakeCBlockClusterAdaptor(CGridDesc_M_N{}, 1, 1, 1));
template <bool HasMainKBlockLoop, bool HasDoubleTailKBlockLoop>
__device__ static void
Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
FloatAB* __restrict__ p_shared_block,
const AGridDesc_B_K0_M0_M1_K1& a_grid_desc_b_k0_m0_m1_k1,
const BGridDesc_B_K0_N0_N1_K1& b_grid_desc_b_k0_n0_n1_k1,
const CGridDesc_M0_M10_M11_N0_N10_N11& c_grid_desc_m0_m10_m11_n0_n10_n11,
const CBlockClusterAdaptor& c_block_cluster_adaptor,
integral_constant<bool, HasMainKBlockLoop>,
integral_constant<bool, HasDoubleTailKBlockLoop>)
{
const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_b_k0_m0_m1_k1.GetElementSpaceSize());
const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_b_k0_n0_n1_k1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_m0_m10_m11_n0_n10_n11.GetElementSpaceSize());
// divide block work by [M, N]
const auto block_work_idx =
c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
const index_t k_batch_id = block_work_idx[I0];
if(!c_block_cluster_adaptor.ValidCTileIndex(
make_tuple(block_work_idx[I1], block_work_idx[I2]),
make_tuple(c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I0),
c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I3))))
{
return;
}
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
const index_t n_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I2]);
// TODO: change this. I think it needs multi-dimensional alignment
constexpr auto max_lds_align = K1;
// TODO: check alignment
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_block_desc_b_k0_m0_m1_k1 = make_naive_tensor_descriptor_aligned(
make_tuple(I1, Number<K0PerBlock>{}, I1, Number<MPerBlock>{}, K1), max_lds_align);
// TODO: check alignment
// B matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto b_block_desc_b_k0_n0_n1_k1 = make_naive_tensor_descriptor_aligned(
make_tuple(I1, Number<K0PerBlock>{}, I1, Number<NPerBlock>{}, K1), max_lds_align);
// TODO: check alignment
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_block_desc_k0_m0_m1_k1 = make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, I1, Number<MPerBlock>{}, K1), max_lds_align);
// TODO: check alignment
// B matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto b_block_desc_k0_n0_n1_k1 = make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, I1, Number<NPerBlock>{}, K1), max_lds_align);
// TODO: check alignment
// A matrix in LDS memory, for blockwise GEMM
constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
// TODO: check alignment
// B matrix in LDS memory, for blockwise GEMM
constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
static_assert(a_block_desc_k0_m0_m1_k1.GetElementSpaceSize() ==
a_k0_m_k1_block_desc.GetElementSpaceSize() &&
b_block_desc_k0_n0_n1_k1.GetElementSpaceSize() ==
b_k0_n_k1_block_desc.GetElementSpaceSize() &&
"wrong!");
// A matrix blockwise copy
auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
BlockSize,
InMemoryDataOperationEnum::Set,
Sequence<1, K0PerBlock, 1, MPerBlock, K1.value>,
ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
ABlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
remove_reference_t<decltype(a_grid_desc_b_k0_m0_m1_k1)>,
decltype(a_block_desc_b_k0_m0_m1_k1),
ABlockTransferSrcAccessOrder,
Sequence<0, 1, 2, 3, 4>,
ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1, // SrcVectorTensorLengths
ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1, // DstVectorTensorLengths
ABlockTransferSrcVectorTensorContiguousDimOrder, // SrcVectorTensorContiguousDimOrder
Sequence<0, 1, 2, 3, 4>, // DstVectorTensorContiguousDimOrder
false,
true>(a_grid_desc_b_k0_m0_m1_k1,
make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0, 0),
a_block_desc_b_k0_m0_m1_k1,
make_multi_index(0, 0, 0, 0, 0));
// B matrix blockwise copy
auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
BlockSize,
InMemoryDataOperationEnum::Set,
Sequence<1, K0PerBlock, 1, NPerBlock, K1.value>,
BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
BBlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
remove_reference_t<decltype(b_grid_desc_b_k0_n0_n1_k1)>,
decltype(b_block_desc_b_k0_n0_n1_k1),
BBlockTransferSrcAccessOrder,
Sequence<0, 1, 2, 3, 4>,
BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1, // SrcVectorTensorLengths
BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1, // DstVectorTensorLengths
BBlockTransferSrcVectorTensorContiguousDimOrder, // SrcVectorTensorContiguousDimOrder
Sequence<0, 1, 2, 3, 4>, // DstVectorTensorContiguousDimOrder
false,
true>(b_grid_desc_b_k0_n0_n1_k1,
make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0, 0),
b_block_desc_b_k0_n0_n1_k1,
make_multi_index(0, 0, 0, 0, 0));
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[K0PerBlock, MPerBlock] is in LDS
// b_mtx[KPerBlocl, NPerBlock] is in LDS
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register
const auto blockwise_gemm =
BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2<
BlockSize,
FloatAB,
FloatAB,
FloatAcc,
decltype(a_k0_m_k1_block_desc),
decltype(b_k0_n_k1_block_desc),
M1PerThreadM111,
N1PerThreadN111,
KPerThread,
M11N11ThreadClusterM110Xs,
M11N11ThreadClusterN110Xs,
M1PerThreadM111,
N1PerThreadN111>{};
constexpr auto c_m10_m11_n10_n11_thread_tensor_lengths =
decltype(blockwise_gemm)::GetCThreadTensorLengths_BM0_BM1_BN0_BN1();
constexpr auto c_thread_desc_m10_m11_n10_n11 = make_naive_tensor_descriptor_packed(
sequence_to_tuple_of_number(c_m10_m11_n10_n11_thread_tensor_lengths));
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
a_block_desc_k0_m0_m1_k1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
b_block_desc_k0_n0_n1_k1.GetElementSpaceSize(), max_lds_align);
FloatAB* p_a_block_double = p_shared_block;
FloatAB* p_b_block_double = p_shared_block + 2 * a_block_aligned_space_size;
// register allocation for output
auto c_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAcc>(
c_thread_desc_m10_m11_n10_n11.GetElementSpaceSize());
// Initialize C
c_thread_buf.Clear();
constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0, 0);
auto a_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_a_block_double, a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
auto b_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_b_block_double, b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
auto a_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_a_block_double + a_block_aligned_space_size,
a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
auto b_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_b_block_double + b_block_aligned_space_size,
b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
// LDS double buffer: preload data into LDS
{
a_blockwise_copy.RunRead(a_grid_desc_b_k0_m0_m1_k1, a_global_buf);
b_blockwise_copy.RunRead(b_grid_desc_b_k0_n0_n1_k1, b_global_buf);
a_blockwise_copy.RunWrite(a_block_desc_b_k0_m0_m1_k1, a_block_even_buf);
b_blockwise_copy.RunWrite(b_block_desc_b_k0_n0_n1_k1, b_block_even_buf);
}
if constexpr(HasMainKBlockLoop)
{
const auto K0 = a_grid_desc_b_k0_m0_m1_k1.GetLength(I1);
index_t k_block_data_begin = 0;
// LDS double buffer: main body
// use Do-While loop instead of For loop to simplify control flow
do
{
// even iteration
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_b_k0_m0_m1_k1,
a_block_slice_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_b_k0_n0_n1_k1,
b_block_slice_copy_step);
// LDS doubel buffer: load next data from device mem
a_blockwise_copy.RunRead(a_grid_desc_b_k0_m0_m1_k1, a_global_buf);
b_blockwise_copy.RunRead(b_grid_desc_b_k0_n0_n1_k1, b_global_buf);
block_sync_lds();
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(c_thread_desc_m10_m11_n10_n11,
a_block_even_buf,
b_block_even_buf,
c_thread_buf);
// LDS double buffer: store next data to LDS
a_blockwise_copy.RunWrite(a_block_desc_b_k0_m0_m1_k1, a_block_odd_buf);
b_blockwise_copy.RunWrite(b_block_desc_b_k0_n0_n1_k1, b_block_odd_buf);
// odd iteration
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_b_k0_m0_m1_k1,
a_block_slice_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_b_k0_n0_n1_k1,
b_block_slice_copy_step);
// LDS doubel buffer: load next data from device mem
a_blockwise_copy.RunRead(a_grid_desc_b_k0_m0_m1_k1, a_global_buf);
b_blockwise_copy.RunRead(b_grid_desc_b_k0_n0_n1_k1, b_global_buf);
block_sync_lds();
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(
c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
// LDS double buffer: store next data to LDS
a_blockwise_copy.RunWrite(a_block_desc_b_k0_m0_m1_k1, a_block_even_buf);
b_blockwise_copy.RunWrite(b_block_desc_b_k0_n0_n1_k1, b_block_even_buf);
k_block_data_begin += 2 * K0PerBlock;
} while(k_block_data_begin < K0 - 2 * K0PerBlock);
}
// LDS double buffer: tail
if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
{
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_b_k0_m0_m1_k1, a_block_slice_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_b_k0_n0_n1_k1, b_block_slice_copy_step);
block_sync_lds();
// LDS double buffer: load last data from device mem
a_blockwise_copy.RunRead(a_grid_desc_b_k0_m0_m1_k1, a_global_buf);
b_blockwise_copy.RunRead(b_grid_desc_b_k0_n0_n1_k1, b_global_buf);
// LDS double buffer: GEMM on 2nd-last data
blockwise_gemm.Run(
c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
// LDS double buffer: store last data to LDS
a_blockwise_copy.RunWrite(a_block_desc_b_k0_m0_m1_k1, a_block_odd_buf);
b_blockwise_copy.RunWrite(b_block_desc_b_k0_n0_n1_k1, b_block_odd_buf);
block_sync_lds();
// LDS double buffer: GEMM on last data
blockwise_gemm.Run(
c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
}
else // if has 1 iteration left
{
__syncthreads();
// LDS double buffer: GEMM on last data
blockwise_gemm.Run(
c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
}
// output: register to global memory
{
constexpr auto c_thread_desc_m0_m10_m11_n0_n10_n11 =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<c_m10_m11_n10_n11_thread_tensor_lengths[I0]>{},
Number<c_m10_m11_n10_n11_thread_tensor_lengths[I1]>{},
I1,
Number<c_m10_m11_n10_n11_thread_tensor_lengths[I2]>{},
Number<c_m10_m11_n10_n11_thread_tensor_lengths[I3]>{}));
const auto c_m10_m11_n10_n11_thread_origin_idx_on_block =
blockwise_gemm.CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
get_thread_local_1d_id());
ThreadwiseTensorSliceTransfer_v1r3<
FloatAcc,
FloatC,
decltype(c_thread_desc_m0_m10_m11_n0_n10_n11),
decltype(c_grid_desc_m0_m10_m11_n0_n10_n11),
ck::tensor_operation::element_wise::PassThrough,
Sequence<1,
c_m10_m11_n10_n11_thread_tensor_lengths[I0],
c_m10_m11_n10_n11_thread_tensor_lengths[I1],
1,
c_m10_m11_n10_n11_thread_tensor_lengths[I2],
c_m10_m11_n10_n11_thread_tensor_lengths[I3]>,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
CGlobalMemoryDataOperation,
1,
true>{c_grid_desc_m0_m10_m11_n0_n10_n11,
make_multi_index(m_block_data_idx_on_grid,
c_m10_m11_n10_n11_thread_origin_idx_on_block[I0],
c_m10_m11_n10_n11_thread_origin_idx_on_block[I1],
n_block_data_idx_on_grid,
c_m10_m11_n10_n11_thread_origin_idx_on_block[I2],
c_m10_m11_n10_n11_thread_origin_idx_on_block[I3]),
ck::tensor_operation::element_wise::PassThrough{}}
.Run(c_thread_desc_m0_m10_m11_n0_n10_n11,
make_tuple(I0, I0, I0, I0, I0, I0),
c_thread_buf,
c_grid_desc_m0_m10_m11_n0_n10_n11,
c_grid_buf);
}
}
};
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp 0 → 100644
View file @ f8a6c69c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/multi_index_transform_helper.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
template <typename GridwiseOp,
typename ADataType,
typename BDataType,
typename DsPointer,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
typename Block2CTileMap,
typename ComputePtrOffsetOfBatch,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle(
const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const index_t batch_count,
const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock_,
const Block2CTileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
__shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
DsPointer p_ds_grid_grp;
static constexpr index_t NumDTensor =
DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
static_for<0, NumDTensor, 1>{}(
[&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset,
p_ds_grid_grp,
p_e_grid + e_batch_offset,
p_shared,
a_grid_desc_k0_m_k1,
b_grid_desc_k0_n_k1,
ds_grid_desc_mblock_mperblock_nblock_nperblock,
e_grid_desc_mblock_mperblock_nblock_nperblock_,
a_element_op,
b_element_op,
cde_element_op,
block_2_ctile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_ds_grid;
ignore = p_e_grid;
ignore = batch_count;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1;
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
ignore = a_element_op;
ignore = b_element_op;
ignore = cde_element_op;
ignore = compute_ptr_offset_of_batch;
ignore = block_2_ctile_map;
#endif
}
template <typename GridwiseOp,
typename ADataType,
typename BDataType,
typename DsPointer,
typename EDataType,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename ComputePtrOffsetOfBatch,
typename Block2CTileMap,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_contraction_multiple_d_wmma_cshuffle(
const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
const index_t batch_count,
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2CTileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
// printf("entry kernel launch");
__shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
// printf("before compute_ptr_offset call");
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
static constexpr index_t NumDTensor =
DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
DsPointer p_ds_grid_grp;
// printf("before allocate pointer d");
static_for<0, NumDTensor, 1>{}(
[&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
// printf("before entry");
GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset,
p_ds_grid_grp,
p_e_grid + e_batch_offset,
p_shared,
a_grid_desc_k0_m_k1,
b_grid_desc_k0_n_k1,
ds_grid_desc_mblock_mperblock_nblock_nperblock,
e_grid_desc_mblock_mperblock_nblock_nperblock,
a_element_op,
b_element_op,
cde_element_op,
block_2_etile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_ds_grid;
ignore = p_e_grid;
ignore = batch_count;
ignore = a_element_op;
ignore = b_element_op;
ignore = cde_element_op;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1;
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = block_2_etile_map;
ignore = compute_ptr_offset_of_batch;
#endif
}
template <typename GridwiseOp,
typename ADataType,
typename BDataType,
typename DsPointer,
typename EDataType,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename Block2CTileMap,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_mupltipe_d_wmma_cshuffle(
const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
__shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid,
p_b_grid,
p_ds_grid,
p_e_grid,
p_shared,
a_grid_desc_k0_m_k1,
b_grid_desc_k0_n_k1,
ds_grid_desc_mblock_mperblock_nblock_nperblock,
e_grid_desc_mblock_mperblock_nblock_nperblock,
a_element_op,
b_element_op,
cde_element_op,
block_2_ctile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_ds_grid;
ignore = p_e_grid;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1;
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = a_element_op;
ignore = b_element_op;
ignore = cde_element_op;
ignore = block_2_ctile_map;
#endif // end of if (defined(__gfx1100__))
}
template < // DataType Family
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
// InMemory Data Descriptor
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename DsGridDesc_M_N,
typename EGridDesc_M_N,
// ElementwiseOp Family
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
InMemoryDataOperationEnum EGlobalMemoryDataOperation,
// Tiling Family
index_t MPerBlock,
index_t NPerBlock,
index_t K0PerBlock,
index_t MPerWmma,
index_t NPerWmma,
index_t K1Value,
index_t MRepeat,
index_t NRepeat,
// ThreadCluster Family
index_t BlockSize,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_K1,
bool AThreadTransferSrcResetCoordinateAfterRun,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_K1,
bool BThreadTransferSrcResetCoordinateAfterRun,
bool BBlockLdsExtraN,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
index_t NumGemmKPrefetchStage = 1,
LoopScheduler LoopSched = make_default_loop_scheduler(),
PipelineVersion PipelineVer = PipelineVersion::v1>
struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
{
static constexpr index_t NumDTensor = DsDataType::Size();
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>{};
static constexpr auto I5 = Number<5>{};
static constexpr auto I6 = Number<6>{};
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
using GridwiseGemmPipe = remove_cvref_t<decltype(
GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
__host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
{
constexpr auto max_lds_align = K1;
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_k0perblock_mperblock_k1 = [&]() {
if constexpr(ABlockLdsExtraM)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
}
}();
return a_block_desc_k0perblock_mperblock_k1;
}
__host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
{
constexpr auto max_lds_align = K1;
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_k0perblock_nperblock_k1 = [&]() {
if constexpr(BBlockLdsExtraN)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
}
}();
return b_block_desc_k0perblock_nperblock_k1;
}
__host__ __device__ static constexpr auto
// *Caution Here repeat is shuffle repeat
GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat()
{
constexpr index_t MWave = MPerBlock / (MRepeat * MPerWmma);
constexpr index_t NWave = NPerBlock / (NRepeat * NPerWmma);
constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMRepeatPerShuffle * MWave * MPerWmma>{},
I1,
Number<CShuffleNRepeatPerShuffle * NWave * NPerWmma>{}));
return c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat;
}
// ck::Tuple<const D0DataType*, const D1DataType*, ...>
static constexpr auto MakeDsGridPointer()
{
return generate_tuple(
[&](auto i) {
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
return static_cast<const DDataType*>(nullptr);
},
Number<NumDTensor>{});
}
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_desc_k0perblock_mperblock_k1 =
GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
constexpr auto b_block_desc_k0perblock_nperblock_k1 =
GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
constexpr auto max_lds_align = K1;
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize(), max_lds_align);
return (a_block_space_size_aligned * sizeof(ADataType) +
b_block_space_size_aligned * sizeof(BDataType));
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
template <typename Block2CTileMap>
__host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
const DsGridDesc_M_N& ds_grid_desc_m_n,
const EGridDesc_M_N& e_grid_desc_m_n,
const Block2CTileMap& block_2_ctile_map)
{
static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
"wrong! K1 need to be known at compile-time");
static_assert((MPerBlock % (MPerWmma * MRepeat) == 0) &&
(NPerBlock % (NRepeat * NPerWmma)) == 0,
"Invalid tuning param!");
const auto M = a_grid_desc_k0_m_k1.GetLength(I1);
const auto N = b_grid_desc_k0_n_k1.GetLength(I1);
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
bool valid = true;
static_for<0, NumDTensor, 1>{}([&](auto i) {
valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
N == ds_grid_desc_m_n[i].GetLength(I1));
});
if(!valid)
{
return false;
}
if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1) &&
K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
return false;
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
return false;
// check gridwise gemm pipeline
const auto num_k_loop = K0 / K0PerBlock;
if(!GridwiseGemmPipe::IsSupported(num_k_loop))
{
return false;
}
if(!block_2_ctile_map.CheckValidity(e_grid_desc_m_n))
{
return false;
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return true;
}
__host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
{
const index_t num_loop = K / (K0PerBlock * K1);
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
}
// E desc for destination in blockwise copy
template <typename EGridDesc_M_N_>
__host__ __device__ static constexpr auto
MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N_& e_grid_desc_m_n)
{
const auto M = e_grid_desc_m_n.GetLength(I0);
const auto N = e_grid_desc_m_n.GetLength(I1);
const auto MBlock = M / MPerBlock;
const auto NBlock = N / NPerBlock;
const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
e_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
return e_grid_desc_mblock_mperblock_nblock_nperblock;
}
// Ds desc for source in blockwise copy
template <typename DsGridDesc_M_N_>
__host__ __device__ static constexpr auto
MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N_& ds_grid_desc_m_n)
{
return generate_tuple(
[&](auto i) {
return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
},
Number<NumDTensor>{});
}
// return block_id to C matrix tile idx (m0, n0) mapping
__host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
const EGridDesc_M_N& e_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
{
return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
e_grid_desc_m_n);
}
using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
using DefaultBlock2CTileMap =
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(EGridDesc_M_N{}, 1, 1))>;
using DsGridPointer = decltype(MakeDsGridPointer());
template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
__device__ static void Run(const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsGridPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
void* __restrict__ p_shared,
const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
e_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op,
const Block2CTileMap& block_2_ctile_map)
{
// printf("safe entry");
// clang-format off
/*******************************************************************************/
// Memory buffer zone.
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
const auto ds_grid_buf = generate_tuple(
[&](auto i) {
return make_dynamic_buffer<AddressSpaceEnum::Global>(
p_ds_grid[i],
ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
},
Number<NumDTensor>{});
auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
/*******************************************************************************/
// BlockIdx.x -> [BlockId.m, BlockId.n]
const auto block_work_idx = block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_ctile_map.ValidCTileIndex(
block_work_idx,
make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{ return; }
// Store BlockId into SGPR
const index_t m_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
const index_t n_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
/*******************************************************************************/
// BlockLevel, A/B Matrix ThreadMapping in LDS, As Destinaion of BlockWise_Copy
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
constexpr auto max_lds_align = K1;
constexpr auto a_block_desc_k0perblock_mperblock_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
constexpr auto b_block_desc_k0perblock_nperblock_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1< ThisThreadBlock,
/* typename SrcElementwiseOperation, */ AElementwiseOperation,
/* typename DstElementwiseOperation, */ ck::tensor_operation::element_wise::PassThrough,
/* InMemoryDataOperationEnum DstInMemOp, */ InMemoryDataOperationEnum::Set,
/* typename BlockSliceLengths, */ Sequence<K0PerBlock, MPerBlock, K1>,
/* typename ThreadClusterLengths, */ ABlockTransferThreadClusterLengths_K0_M_K1,
/* typename ThreadClusterArrangeOrder, */ ABlockTransferThreadClusterArrangeOrder,
/* typename SrcData, */ ADataType,
/* typename DstData, */ ADataType,
/* typename SrcDesc, */ decltype(a_grid_desc_k0_m_k1),
/* typename DstDesc, */ decltype(a_block_desc_k0perblock_mperblock_k1),
/* typename SrcDimAccessOrder, */ ABlockTransferSrcAccessOrder,
/* typename DstDimAccessOrder, */ Sequence<0, 1, 2>,
/* index_t SrcVectorDim, */ ABlockTransferSrcVectorDim,
/* index_t DstVectorDim, */ 2,
/* index_t SrcScalarPerVector, */ ABlockTransferSrcScalarPerVector,
/* index_t DstScalarPerVector, */ ABlockTransferDstScalarPerVector_K1,
/* index_t SrcScalarStrideInVector, */ 1,
/* index_t DstScalarStrideInVector, */ 1,
/* bool ThreadTransferSrcResetCoordinateAfterRun, */ AThreadTransferSrcResetCoordinateAfterRun,
/* bool ThreadTransferDstResetCoordinateAfterRun, */ true>(
a_grid_desc_k0_m_k1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_k0perblock_mperblock_k1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BDataType,
BDataType,
decltype(b_grid_desc_k0_n_k1),
decltype(b_block_desc_k0perblock_nperblock_k1),
BBlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
BBlockTransferSrcVectorDim,
2,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true>(
b_grid_desc_k0_n_k1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_k0perblock_nperblock_k1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
/*******************************************************************************/
// GEMM
constexpr auto WmmaK = 16;
constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
auto blockwise_gemm =
BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO<BlockSize,
ADataType,
BDataType,
AccDataType,
decltype(a_block_desc_k0perblock_mperblock_k1),
decltype(b_block_desc_k0perblock_nperblock_k1),
MPerWmma,
NPerWmma,
MRepeat,
NRepeat,
KPack>{};
// Prepare Register for C matrix
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
/*******************************************************************************/
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
// LDS allocation for A and B: be careful of alignment
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<ADataType*>(p_shared), a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<BDataType*>(p_shared) + a_block_space_size_aligned, b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize());
// Shift Per SUB_K
constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
// gridwise GEMM pipeline
const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
a_block_desc_k0perblock_mperblock_k1,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_grid_desc_k0_n_k1,
b_block_desc_k0perblock_nperblock_k1,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
blockwise_gemm,
c_thread_buf,
K0BlockMainLoop);
/*******************************************************************************/
//printf("safe 1");
// write out to C, implement shuffle
{
constexpr auto c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs =
blockwise_gemm.GetCThreadDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
// This API Provide All dimension (size) you need
constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp =
blockwise_gemm.GetCBlockDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
constexpr auto MWave = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I1);
constexpr auto MSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I2);
constexpr auto NWave = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I4);
constexpr auto NThreadPerSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I5);
constexpr auto MAccVgprs = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I6);
// LDS descriptor, shuffle and write out in MRepeat x NRepeat times
constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<CShuffleDataType*>(p_shared),
c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat.GetElementSpaceSize());
constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs = transform_tensor_descriptor(
c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMRepeatPerShuffle>{}, // MRepeat per shuffle repeat
MWave, // MWave
MSubGroup, // MSubGroup * MAccVgprs = MPerWmma
MAccVgprs)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNRepeatPerShuffle>{}, // NRepeat per shuffle repeat
NWave, // NWave
NThreadPerSubGroup))), // NThreadPerSubGroup = NPerWmma
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<>{}, Sequence<0, 1, 2, 6>{}, Sequence<>{}, Sequence<3, 4, 5>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block = blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(MRepeat, MWave, MSubGroup, MAccVgprs))),
make_tuple(Sequence<0, 1, 2, 3>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(NRepeat, NWave, NThreadPerSubGroup))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx = m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_idx = n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
CShuffleDataType,
decltype(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
decltype(c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMRepeatPerShuffle,
I1,
I1,
CShuffleNRepeatPerShuffle,
I1,
I1,
MAccVgprs>,
Sequence<0, 1, 2, 3, 4, 5, 6>,
6,
1, // vector write pixel
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
make_multi_index(0,
m_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
0,
n_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3]),
ck::tensor_operation::element_wise::PassThrough{}};
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat),
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor buffers
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin = container_concat(
make_tuple(make_multi_index(0, 0, 0, 0)),
generate_tuple(
[&](auto) {
return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
},
Number<NumDTensor>{}));
// shuffle: blockwise copy C from LDS to global
auto cde_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v7<
ThisThreadBlock, // ThreadGroup
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CDEElementwiseOperation, // ElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // DstInMemOp,
Sequence<1,
CShuffleMRepeatPerShuffle * MWave * MPerWmma,
1,
CShuffleNRepeatPerShuffle * NWave * NPerWmma>, // BlockSliceLengths,
CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CDEShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // bool ThreadTransferSrcResetCoordinateAfterRun,
Sequence<false>> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
cde_element_op};
// space filling curve for local reg & global memory
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MRepeat, 1, 1, NRepeat, 1, 1, MAccVgprs>,
Sequence<0, 1, 2, 3, 4, 5, 6>,
Sequence<CShuffleMRepeatPerShuffle,
1,
1,
CShuffleNRepeatPerShuffle,
1,
1,
MAccVgprs>>{};
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_cde_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMRepeatPerShuffle * MWave * MPerWmma,
1,
CShuffleNRepeatPerShuffle * NWave * NPerWmma>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_cde_global.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
cde_shuffle_block_copy_lds_to_global.Run(
c_ds_desc_refs,
c_ds_buf_refs,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
tie(e_grid_buf));
if constexpr(access_id < num_access - 1)
{
constexpr auto cde_global_step = sfc_cde_global.GetForwardStep(access_id);
// move on Ds
static_for<0, NumDTensor, 1>{}([&](auto i) {
cde_shuffle_block_copy_lds_to_global.MoveSrcSliceWindow(
c_ds_desc_refs, i + I1, cde_global_step);
});
// move on E
cde_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
I0,
cde_global_step);
}
});
}
// clang-format on
}
};
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_normalization_naive_variance.hpp
View file @ f8a6c69c
......@@ -4,9 +4,8 @@
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/reduction_common.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/utility/reduction_functions_accumulate.hpp"
#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
......@@ -19,8 +18,8 @@ template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
typename AccDataType,
typename AccElementwiseOperation,
typename ComputeDataType,
typename YElementwiseOperation,
typename GridDesc_M_K,
index_t BlockSize,
index_t MThreadClusterSize,
......@@ -46,6 +45,10 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
(YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
"Invalid thread slice sizes and/or vector sizes configuration, please check!");
static_assert(XSrcVectorSize == YDstVectorSize);
static_assert(XSrcVectorSize == GammaSrcVectorSize);
static_assert(XSrcVectorSize == BetaSrcVectorSize);
static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
......@@ -59,19 +62,23 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
static constexpr auto thread_cluster_desc =
make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, XSrcVectorSize>;
static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
using ThreadReduceDstDesc_M =
decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
using BlockwiseSumReduce = PartitionedBlockwiseReduction<AccDataType,
using BlockwiseSumReduce = PartitionedBlockwiseReduction<ComputeDataType,
BlockSize,
ThreadClusterLengths_M_K,
ThreadClusterArrangeOrder,
reduce::Add,
true>;
using ThreadwiseSumReduce = ThreadwiseReduction<AccDataType,
using ThreadwiseSumReduce = ThreadwiseReduction<ComputeDataType,
ThreadReduceSrcDesc_M_K,
ThreadReduceDstDesc_M,
reduce::Add,
......@@ -81,64 +88,70 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
__device__ static void Run(const GridDesc_M_K& x_grid_desc_m_k,
const GridDesc_M_K& gamma_grid_desc_m_k,
const GridDesc_M_K& beta_grid_desc_m_k,
const GridDesc_M_K& y_grid_desc_m_k,
index_t num_k_block_tile_iteration,
AccDataType epsilon,
ComputeDataType epsilon,
const XDataType* const __restrict__ p_x_global,
const GammaDataType* const __restrict__ p_gamma_global,
const BetaDataType* const __restrict__ p_beta_global,
YDataType* const __restrict__ p_y_global,
const AccElementwiseOperation acc_elementwise_op)
const YElementwiseOperation y_elementwise_op)
{
if constexpr(SweepOnce)
{
num_k_block_tile_iteration = 1;
}
// LDS
__shared__ AccDataType p_reduce_work_buffer[BlockSize];
auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
__shared__ ComputeDataType p_reduce_work_buffer[BlockSize];
auto reduce_work_buf =
make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
x_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
gamma_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
MThreadSliceSize * KThreadSliceSize,
true>& beta_thread_buf = gamma_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
y_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
MThreadSliceSize * KThreadSliceSize,
true>& x_square_thread_buf = y_thread_buf;
auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
mean_square_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>& var_thread_buf =
auto x_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
ComputeDataType,
MThreadSliceSize * XSrcVectorSize,
true>{};
},
Number<ThreadBufferNumber>{});
auto gamma_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
ComputeDataType,
MThreadSliceSize * GammaSrcVectorSize,
true>{};
},
Number<ThreadBufferNumber>{});
auto& beta_thread_buf = gamma_thread_buf;
auto y_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
ComputeDataType,
MThreadSliceSize * YDstVectorSize,
true>{};
},
Number<ThreadBufferNumber>{});
auto& x_square_thread_buf = y_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
mean_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
mean_square_thread_buf;
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
mean_thread_buf(I) = reduce::Add::template GetIdentityValue<AccDataType>();
mean_square_thread_buf(I) = reduce::Add::template GetIdentityValue<AccDataType>();
});
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>&
var_thread_buf = mean_square_thread_buf;
const index_t thread_local_id = get_thread_local_1d_id();
const index_t block_global_id = get_block_1d_id();
......@@ -149,12 +162,8 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
const auto thread_m_cluster_id = thread_cluster_idx[I0];
const auto thread_k_cluster_id = thread_cluster_idx[I1];
using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, KThreadSliceSize>;
constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
AccDataType,
ComputeDataType,
GridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
......@@ -166,11 +175,11 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
x_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize));
thread_k_cluster_id * XSrcVectorSize));
auto threadwise_gamma_load =
ThreadwiseTensorSliceTransfer_v2<GammaDataType,
AccDataType,
ComputeDataType,
GridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
......@@ -182,11 +191,11 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
gamma_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize));
thread_k_cluster_id * GammaSrcVectorSize));
auto threadwise_beta_load =
ThreadwiseTensorSliceTransfer_v2<BetaDataType,
AccDataType,
ComputeDataType,
GridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
......@@ -198,14 +207,14 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
beta_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize));
thread_k_cluster_id * BetaSrcVectorSize));
auto threadwise_y_store =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
YDataType,
decltype(thread_buffer_desc_m_k),
GridDesc_M_K,
AccElementwiseOperation,
YElementwiseOperation,
ThreadBufferLengths_M_K,
ThreadBufferDimAccessOrder,
YDstVectorDim,
......@@ -216,13 +225,10 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
y_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize),
acc_elementwise_op);
thread_k_cluster_id * YDstVectorSize),
y_elementwise_op);
// Copy x from Cache
// one pass: fwd, second pass: bwd
constexpr auto thread_copy_fwd_step_m_k =
make_multi_index(0, SweepOnce ? 0 : K_BlockTileSize);
constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
constexpr auto thread_copy_bwd_step_m_k =
make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
......@@ -239,121 +245,260 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
// FIXME: Should not hack the transform from deviceOP
int reduce_length = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
index_t reducedTiles = 0;
do
{
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf);
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
mean_thread_buf(I) = reduce::Add::template GetIdentityValue<ComputeDataType>();
mean_square_thread_buf(I) = reduce::Add::template GetIdentityValue<ComputeDataType>();
});
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
x_square_thread_buf(Number<offset_m_k>{}) =
x_thread_buf(Number<offset_m_k>{}) * x_thread_buf(Number<offset_m_k>{});
// Separate sweep once and sweep twice pipeline
if constexpr(SweepOnce)
{
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf(i));
threadwise_gamma_load.Run(gamma_grid_desc_m_k,
gamma_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
gamma_thread_buf(i));
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
x_square_thread_buf(i)(Number<offset_m_k>{}) =
x_thread_buf(i)(Number<offset_m_k>{}) *
x_thread_buf(i)(Number<offset_m_k>{});
});
});
});
ThreadwiseSumReduce::Reduce(x_thread_buf, mean_thread_buf);
ThreadwiseSumReduce::Reduce(x_square_thread_buf, mean_square_thread_buf);
ThreadwiseSumReduce::Reduce(x_thread_buf[i], mean_thread_buf);
ThreadwiseSumReduce::Reduce(x_square_thread_buf[i], mean_square_thread_buf);
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
if constexpr(i != ThreadBufferNumber - 1)
{
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
thread_copy_fwd_step_m_k);
}
});
++reducedTiles;
} while(reducedTiles < num_k_block_tile_iteration);
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
if constexpr(I > 0)
block_sync_lds();
BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
if constexpr(I > 0)
block_sync_lds();
BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
// var(x) = E[x^2] - E[x]^2
var_thread_buf(I) =
mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
});
block_sync_lds();
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
// normalize
y_thread_buf(iK0)(Number<offset_m_k>{}) =
(x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
divisor;
// gamma & beta
y_thread_buf(iK0)(Number<offset_m_k>{}) =
y_thread_buf(iK0)(Number<offset_m_k>{}) *
gamma_thread_buf(iK0)(Number<offset_m_k>{});
});
});
});
BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_beta_load.Run(beta_grid_desc_m_k,
beta_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
beta_thread_buf(i));
// var(x) = E[x^2] - E[x]^2
var_thread_buf(I) =
mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
});
if constexpr(i != ThreadBufferNumber - 1)
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
// y = (x - E[x]) / sqrt(var[x] + epsilon)
auto thread_copy_tail_m_k = (num_k_block_tile_iteration - 1) * thread_copy_fwd_step_m_k;
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
// beta
y_thread_buf(iK0)(Number<offset_m_k>{}) =
y_thread_buf(iK0)(Number<offset_m_k>{}) +
beta_thread_buf(iK0)(Number<offset_m_k>{});
});
});
});
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_y_store.Run(thread_buffer_desc_m_k,
make_tuple(I0, I0),
y_thread_buf(i),
y_grid_desc_m_k,
y_global_val_buf);
reducedTiles = 0;
do
if constexpr(i != ThreadBufferNumber - 1)
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
} // end of sweep once
else
{
if constexpr(!SweepOnce)
for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
{
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf);
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf(i));
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
x_square_thread_buf(i)(Number<offset_m_k>{}) =
x_thread_buf(i)(Number<offset_m_k>{}) *
x_thread_buf(i)(Number<offset_m_k>{});
});
});
ThreadwiseSumReduce::Reduce(x_thread_buf[i], mean_thread_buf);
ThreadwiseSumReduce::Reduce(x_square_thread_buf[i], mean_square_thread_buf);
});
}
threadwise_gamma_load.Run(gamma_grid_desc_m_k,
gamma_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
gamma_thread_buf);
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
if constexpr(I > 0)
block_sync_lds();
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
// normalize
y_thread_buf(Number<offset_m_k>{}) =
(x_thread_buf(Number<offset_m_k>{}) - mean_thread_buf(iM)) /
sqrt(var_thread_buf(iM) + epsilon);
// gamma
y_thread_buf(Number<offset_m_k>{}) =
y_thread_buf(Number<offset_m_k>{}) * gamma_thread_buf(Number<offset_m_k>{});
});
});
BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
threadwise_beta_load.Run(beta_grid_desc_m_k,
beta_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
beta_thread_buf);
block_sync_lds();
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
// beta
y_thread_buf(Number<offset_m_k>{}) =
y_thread_buf(Number<offset_m_k>{}) + beta_thread_buf(Number<offset_m_k>{});
});
// var(x) = E[x^2] - E[x]^2
var_thread_buf(I) =
mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
});
threadwise_y_store.Run(thread_buffer_desc_m_k,
make_tuple(I0, I0),
y_thread_buf,
y_grid_desc_m_k,
y_global_val_buf);
auto thread_copy_tail_m_k =
(num_k_block_tile_iteration - 1) * ThreadBufferNumber * thread_copy_fwd_step_m_k;
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
{
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf(i));
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
});
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_gamma_load.Run(gamma_grid_desc_m_k,
gamma_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
gamma_thread_buf(i));
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
++reducedTiles;
} while(reducedTiles < num_k_block_tile_iteration);
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
// normalize
y_thread_buf(iK0)(Number<offset_m_k>{}) =
(x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
divisor;
// gamma
y_thread_buf(iK0)(Number<offset_m_k>{}) =
y_thread_buf(iK0)(Number<offset_m_k>{}) *
gamma_thread_buf(iK0)(Number<offset_m_k>{});
});
});
});
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_beta_load.Run(beta_grid_desc_m_k,
beta_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
beta_thread_buf(i));
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
// beta
y_thread_buf(iK0)(Number<offset_m_k>{}) =
y_thread_buf(iK0)(Number<offset_m_k>{}) +
beta_thread_buf(iK0)(Number<offset_m_k>{});
});
});
});
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_y_store.Run(thread_buffer_desc_m_k,
make_tuple(I0, I0),
y_thread_buf(i),
y_grid_desc_m_k,
y_global_val_buf);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
2 * thread_copy_bwd_step_m_k);
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
2 * thread_copy_bwd_step_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
2 * thread_copy_bwd_step_m_k);
}
} // end of sweep twice
}
};
......
include/ck/tensor_operation/gpu/grid/gridwise_normalization_selector.hpp 0 → 100644
View file @ f8a6c69c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_operation/gpu/grid/gridwise_normalization_naive_variance.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp"
namespace ck {
template <typename GridwiseReduction,
typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
typename ComputeDataType,
typename YElementwiseOperation,
typename GridDesc_M_K>
__global__ void kernel_normalization(const GridDesc_M_K x_grid_desc_m_k,
const GridDesc_M_K gamma_grid_desc_m_k,
const GridDesc_M_K beta_grid_desc_m_k,
const GridDesc_M_K y_grid_desc_m_k,
index_t num_k_block_tile_iteration,
ComputeDataType epsilon,
const XDataType* const __restrict__ p_x_global,
const GammaDataType* const __restrict__ p_gamma_global,
const BetaDataType* const __restrict__ p_beta_global,
YDataType* const __restrict__ p_y_global,
const YElementwiseOperation y_elementwise_op)
{
GridwiseReduction::Run(x_grid_desc_m_k,
gamma_grid_desc_m_k,
beta_grid_desc_m_k,
y_grid_desc_m_k,
num_k_block_tile_iteration,
epsilon,
p_x_global,
p_gamma_global,
p_beta_global,
p_y_global,
y_elementwise_op);
};
template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
typename ComputeDataType,
typename YElementwiseOperation,
typename GridDesc_M_K,
index_t BlockSize,
index_t MThreadClusterSize,
index_t KThreadClusterSize,
index_t MThreadSliceSize,
index_t KThreadSliceSize,
index_t XSrcVectorDim,
index_t XSrcVectorSize,
index_t GammaSrcVectorDim,
index_t GammaSrcVectorSize,
index_t BetaSrcVectorDim,
index_t BetaSrcVectorSize,
index_t YDstVectorDim,
index_t YDstVectorSize,
bool UseWelford>
auto NormalizationKernelSelector(bool isSweepOnce)
{
using GridwiseNormalizationGenericNaive =
GridwiseNormalizationNaiveVariance_mk_to_mk<XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
YElementwiseOperation,
GridDesc_M_K,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
XSrcVectorDim,
XSrcVectorSize,
GammaSrcVectorDim,
GammaSrcVectorSize,
BetaSrcVectorDim,
BetaSrcVectorSize,
YDstVectorDim,
YDstVectorSize,
false>;
using GridwiseNormalizationSweepOnceNaive =
GridwiseNormalizationNaiveVariance_mk_to_mk<XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
YElementwiseOperation,
GridDesc_M_K,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
XSrcVectorDim,
XSrcVectorSize,
GammaSrcVectorDim,
GammaSrcVectorSize,
BetaSrcVectorDim,
BetaSrcVectorSize,
YDstVectorDim,
YDstVectorSize,
true>;
using GridwiseNormalizationGenericWelford =
GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
YElementwiseOperation,
GridDesc_M_K,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
XSrcVectorDim,
XSrcVectorSize,
GammaSrcVectorDim,
GammaSrcVectorSize,
BetaSrcVectorDim,
BetaSrcVectorSize,
YDstVectorDim,
YDstVectorSize,
false>;
using GridwiseNormalizationSweepOnceWelford =
GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
YElementwiseOperation,
GridDesc_M_K,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
XSrcVectorDim,
XSrcVectorSize,
GammaSrcVectorDim,
GammaSrcVectorSize,
BetaSrcVectorDim,
BetaSrcVectorSize,
YDstVectorDim,
YDstVectorSize,
true>;
if constexpr(UseWelford)
{
return isSweepOnce ? kernel_normalization<GridwiseNormalizationSweepOnceWelford,
XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
YElementwiseOperation,
GridDesc_M_K>
: kernel_normalization<GridwiseNormalizationGenericWelford,
XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
YElementwiseOperation,
GridDesc_M_K>;
}
else
{
return isSweepOnce ? kernel_normalization<GridwiseNormalizationSweepOnceNaive,
XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
YElementwiseOperation,
GridDesc_M_K>
: kernel_normalization<GridwiseNormalizationGenericNaive,
XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
YElementwiseOperation,
GridDesc_M_K>;
}
}
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp
View file @ f8a6c69c
......@@ -16,8 +16,8 @@ template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
typename AccDataType,
typename AccElementwiseOperation,
typename ComputeDataType,
typename YElementwiseOperation,
typename GridDesc_M_K,
index_t BlockSize,
index_t MThreadClusterSize,
......@@ -43,6 +43,10 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
(YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
"Invalid thread slice sizes and/or vector sizes configuration, please check!");
static_assert(XSrcVectorSize == YDstVectorSize);
static_assert(XSrcVectorSize == GammaSrcVectorSize);
static_assert(XSrcVectorSize == BetaSrcVectorSize);
static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
......@@ -56,15 +60,19 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
static constexpr auto thread_cluster_desc =
make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, XSrcVectorSize>;
static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
using ThreadReduceDstDesc_M =
decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
using ThreadwiseWelford =
ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
ThreadwiseWelford<ComputeDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
using BlockwiseWelford = BlockwiseWelford<AccDataType,
using BlockwiseWelford = BlockwiseWelford<ComputeDataType,
BlockSize,
ThreadClusterLengths_M_K,
ThreadClusterArrangeOrder>;
......@@ -77,10 +85,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
static constexpr auto XThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
static constexpr auto GammaThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
static constexpr auto BetaThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
static constexpr auto YThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
__device__ static int GetKPerThread(const GridDesc_M_K& x_grid_desc_m_k,
int thread_k_cluster_id)
......@@ -93,7 +98,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
if(kPerBlockTail > 0)
{
static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
int thread_max_len =
(thread_k_cluster_id + 1) * XSrcVectorSize + K_BlockTileStepSize * i;
int delta = thread_max_len - kPerBlockTail;
......@@ -110,59 +115,41 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
const GridDesc_M_K& beta_grid_desc_m_k,
const GridDesc_M_K& y_grid_desc_m_k,
index_t num_k_block_tile_iteration,
AccDataType epsilon,
ComputeDataType epsilon,
const XDataType* const __restrict__ p_x_global,
const GammaDataType* const __restrict__ p_gamma_global,
const BetaDataType* const __restrict__ p_beta_global,
YDataType* const __restrict__ p_y_global,
const AccElementwiseOperation acc_elementwise_op)
const YElementwiseOperation y_elementwise_op)
{
if constexpr(SweepOnce)
{
num_k_block_tile_iteration = 1;
}
auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
auto x_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
ComputeDataType,
MThreadSliceSize * XSrcVectorSize,
true>{};
},
Number<XThreadBufferNumber>{});
Number<ThreadBufferNumber>{});
auto gamma_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
ComputeDataType,
MThreadSliceSize * GammaSrcVectorSize,
true>{};
},
Number<GammaThreadBufferNumber>{});
auto beta_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
MThreadSliceSize * BetaSrcVectorSize,
true>{};
},
Number<BetaThreadBufferNumber>{});
Number<ThreadBufferNumber>{});
auto y_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
MThreadSliceSize * YDstVectorSize,
true>{};
},
Number<YThreadBufferNumber>{});
auto& beta_thread_buf = gamma_thread_buf;
auto& y_thread_buf = x_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
mean_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
var_thread_buf;
const index_t thread_local_id = get_thread_local_1d_id();
const index_t block_global_id = get_block_1d_id();
......@@ -173,12 +160,8 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
const auto thread_m_cluster_id = thread_cluster_idx[I0];
const auto thread_k_cluster_id = thread_cluster_idx[I1];
using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, XSrcVectorSize>;
constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
AccDataType,
ComputeDataType,
GridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
......@@ -194,7 +177,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
auto threadwise_gamma_load =
ThreadwiseTensorSliceTransfer_v2<GammaDataType,
AccDataType,
ComputeDataType,
GridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
......@@ -210,7 +193,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
auto threadwise_beta_load =
ThreadwiseTensorSliceTransfer_v2<BetaDataType,
AccDataType,
ComputeDataType,
GridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
......@@ -225,11 +208,11 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
thread_k_cluster_id * BetaSrcVectorSize));
auto threadwise_y_store =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
YDataType,
decltype(thread_buffer_desc_m_k),
GridDesc_M_K,
AccElementwiseOperation,
YElementwiseOperation,
ThreadBufferLengths_M_K,
ThreadBufferDimAccessOrder,
YDstVectorDim,
......@@ -241,7 +224,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * YDstVectorSize),
acc_elementwise_op);
y_elementwise_op);
constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
constexpr auto thread_copy_bwd_step_m_k =
......@@ -260,67 +243,47 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
threadwise_welford.max_count_ = GetKPerThread(x_grid_desc_m_k, thread_k_cluster_id);
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
var_thread_buf(I) = type_convert<AccDataType>(0.0f);
mean_thread_buf(I) = type_convert<ComputeDataType>(0.0f);
var_thread_buf(I) = type_convert<ComputeDataType>(0.0f);
});
for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
// Separate sweep once and sweep twice pipeline
if constexpr(SweepOnce)
{
static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf(i));
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
});
}
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
if constexpr(I > 0)
block_sync_lds();
int count = threadwise_welford.cur_count_;
BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
});
auto thread_copy_tail_m_k =
(num_k_block_tile_iteration - 1) * XThreadBufferNumber * thread_copy_fwd_step_m_k;
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
{
if constexpr(!SweepOnce)
{
static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf(i));
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
});
}
static_for<0, GammaThreadBufferNumber, 1>{}([&](auto i) {
threadwise_gamma_load.Run(gamma_grid_desc_m_k,
gamma_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
gamma_thread_buf(i));
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
thread_copy_fwd_step_m_k);
threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
if constexpr(i != ThreadBufferNumber - 1)
{
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
thread_copy_fwd_step_m_k);
}
});
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
if constexpr(I > 0)
block_sync_lds();
int count = threadwise_welford.cur_count_;
BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
});
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
......@@ -330,7 +293,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
(x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
divisor;
// gamma
// gamma & beta
y_thread_buf(iK0)(Number<offset_m_k>{}) =
y_thread_buf(iK0)(Number<offset_m_k>{}) *
gamma_thread_buf(iK0)(Number<offset_m_k>{});
......@@ -338,18 +301,20 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
});
});
static_for<0, BetaThreadBufferNumber, 1>{}([&](auto i) {
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_beta_load.Run(beta_grid_desc_m_k,
beta_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
beta_thread_buf(i));
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
thread_copy_fwd_step_m_k);
if constexpr(i != ThreadBufferNumber - 1)
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
......@@ -362,22 +327,134 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
});
});
static_for<0, YThreadBufferNumber, 1>{}([&](auto i) {
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_y_store.Run(thread_buffer_desc_m_k,
make_tuple(I0, I0),
y_thread_buf(i),
y_grid_desc_m_k,
y_global_val_buf);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_fwd_step_m_k);
if constexpr(i != ThreadBufferNumber - 1)
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
} // end of sweep once
else
{
for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
{
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf(i));
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
});
}
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
if constexpr(I > 0)
block_sync_lds();
int count = threadwise_welford.cur_count_;
BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
});
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
2 * thread_copy_bwd_step_m_k);
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
2 * thread_copy_bwd_step_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
}
auto thread_copy_tail_m_k =
(num_k_block_tile_iteration - 1) * ThreadBufferNumber * thread_copy_fwd_step_m_k;
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
{
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf(i));
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
});
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_gamma_load.Run(gamma_grid_desc_m_k,
gamma_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
gamma_thread_buf(i));
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
// normalize
y_thread_buf(iK0)(Number<offset_m_k>{}) =
(x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
divisor;
// gamma
y_thread_buf(iK0)(Number<offset_m_k>{}) =
y_thread_buf(iK0)(Number<offset_m_k>{}) *
gamma_thread_buf(iK0)(Number<offset_m_k>{});
});
});
});
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_beta_load.Run(beta_grid_desc_m_k,
beta_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
beta_thread_buf(i));
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
// beta
y_thread_buf(iK0)(Number<offset_m_k>{}) =
y_thread_buf(iK0)(Number<offset_m_k>{}) +
beta_thread_buf(iK0)(Number<offset_m_k>{});
});
});
});
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_y_store.Run(thread_buffer_desc_m_k,
make_tuple(I0, I0),
y_thread_buf(i),
y_grid_desc_m_k,
y_global_val_buf);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
2 * thread_copy_bwd_step_m_k);
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
2 * thread_copy_bwd_step_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
2 * thread_copy_bwd_step_m_k);
}
} // end of sweep twice
}
};
......
include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp
View file @ f8a6c69c
......@@ -185,9 +185,7 @@ struct GridwiseSparseEmbeddingsForwardLayernorm
},
Number<NumEmbeddings>{});
auto out_data_refs = generate_tie(
[&](auto output_index_) -> auto& {
return acc_thread_buf(Number<register_offset>{});
},
[&](auto) -> auto& { return acc_thread_buf(Number<register_offset>{}); },
Number<1>{});
unpack2(emb_elementwise_op, out_data_refs, in_data_refs);
});
......
include/ck/utility/math_v2.hpp
View file @ f8a6c69c
......@@ -83,6 +83,11 @@ static inline __host__ bool isnan(int4_t x)
};
#endif
static inline __host__ half_t sqrt(half_t x)
{
return static_cast<half_t>(std::sqrt(static_cast<float>(x)));
};
static inline __host__ float sqrt(float x) { return std::sqrt(x); };
static inline __host__ double sqrt(double x) { return std::sqrt(x); };
......@@ -158,9 +163,14 @@ static inline __device__ bool isnan(half_t x)
return (xx & 0x7FFF) > 0x7C00;
};
static inline __device__ float sqrt(float x) { return ::sqrtf(x); };
static inline __device__ half_t sqrt(half_t x)
{
return static_cast<half_t>(__builtin_amdgcn_sqrtf(static_cast<float>(x)));
};
static inline __device__ float sqrt(float x) { return __builtin_amdgcn_sqrtf(x); };
static inline __device__ double sqrt(double x) { return ::sqrt(x); };
static inline __device__ double sqrt(double x) { return __builtin_amdgcn_sqrt(x); };
} // namespace math
} // namespace ck
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View file @ f8a6c69c
......@@ -89,8 +89,10 @@ using Scale = ck::tensor_operation::element_wise::Scale;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using AddReluAdd = ck::tensor_operation::element_wise::AddReluAdd;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
using ScaleAdd = ck::tensor_operation::element_wise::ScaleAdd;
template <typename Activation>
using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;
......
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp 0 → 100644
View file @ f8a6c69c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
F16,
F16,
F16,
F16,
ck::Tuple<F16>,
ck::Tuple<>,
PassThrough,
PassThrough,
ScaleAdd,
PassThrough,
PassThrough,
MaskingSpecialization::MaskOutUpperTriangle>>>&
instances);
void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
std::vector<
std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
F16,
F16,
F16,
F16,
ck::Tuple<F16>,
ck::Tuple<>,
PassThrough,
PassThrough,
ScaleAdd,
PassThrough,
PassThrough,
MaskingSpecialization::MaskDisabled>>>&
instances);
void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
BF16,
BF16,
BF16,
BF16,
ck::Tuple<BF16>,
ck::Tuple<>,
PassThrough,
PassThrough,
ScaleAdd,
PassThrough,
PassThrough,
MaskingSpecialization::MaskOutUpperTriangle>>>&
instances);
void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
std::vector<
std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
BF16,
BF16,
BF16,
BF16,
ck::Tuple<BF16>,
ck::Tuple<>,
PassThrough,
PassThrough,
ScaleAdd,
PassThrough,
PassThrough,
MaskingSpecialization::MaskDisabled>>>&
instances);
template <typename ADataType,
typename B0DataType,
typename B1DataType,
typename CDataType,
typename Acc0BiasDataType,
MaskingSpecialization MaskingSpec>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
ADataType,
B0DataType,
B1DataType,
CDataType,
Acc0BiasDataType,
ck::Tuple<>,
PassThrough,
PassThrough,
ScaleAdd,
PassThrough,
PassThrough,
MaskingSpec>>
{
using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
ADataType,
B0DataType,
B1DataType,
CDataType,
Acc0BiasDataType,
ck::Tuple<>,
PassThrough,
PassThrough,
ScaleAdd,
PassThrough,
PassThrough,
MaskingSpec>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t> &&
Acc0BiasDataType::Size() == 1 &&
is_same_v<tuple_element_t<0, Acc0BiasDataType>, half_t>)
{
if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
{
add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
op_ptrs);
}
else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
{
add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16> &&
Acc0BiasDataType::Size() == 1 &&
is_same_v<tuple_element_t<0, Acc0BiasDataType>, BF16>)
{
if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
{
add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
op_ptrs);
}
else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
{
add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
op_ptrs);
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
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