Skip to content

GitLab

Commit d3341a67 authored by Jing Zhang's avatar Jing Zhang
Browse files

xdlops refactor

parent b62bf8c3
Show whitespace changes
Inline Side-by-side
Showing with 354 additions and 617 deletions
composable_kernel/include/tensor_operation/blockwise_gemm_xdlops.hpp
View file @ d3341a67
......@@ -9,82 +9,99 @@ namespace ck {
template <index_t BlockSize,
typename FloatAB,
class ABlockDesc,
class BBlockDesc,
index_t MPerWave,
index_t NPerWave,
typename AK0MK1BlockDesc,
typename BK0NK1BlockDesc,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t K1>
struct BlockwiseGemmXdlops_km_kn_m0m1m2n_v1
struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
{
using CIndex = MultiIndex<2>;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
using CIndex = MultiIndex<2>;
static constexpr index_t WaveSize = 64;
static constexpr index_t M0 = ABlockDesc{}.GetLength(I1);
static constexpr index_t M1 = ABlockDesc{}.GetLength(I2);
static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
static constexpr index_t K0 = BK0NK1BlockDesc{}.GetLength(I0);
static constexpr index_t KPerBlock = K0;
static constexpr index_t KPack = K1;
static constexpr index_t N0 = BBlockDesc{}.GetLength(I1);
static constexpr index_t N1 = BBlockDesc{}.GetLength(I2);
static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, K1>{};
static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerWave, NPerWave, K1>{};
static constexpr auto CXdlopsLayout = xdlops_gemm.GetCXdlopsLayout();
static constexpr index_t MWaves = M1 / MPerWave;
static constexpr index_t NWaves = N1 / NPerWave;
static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
static constexpr index_t MRepeat = M0;
static constexpr index_t NRepeat = N0;
__device__ static constexpr auto GetCM0N0M1N1M2M3M4N2ThreadDesc()
{
constexpr auto M0 = Number<CXdlopsLayout.M1()>{};
constexpr auto M2 = Number<CXdlopsLayout.M0()>{};
return make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, I1, M0, I1, M2, I1));
}
__device__ constexpr auto GetCLayout() const { return xdlops_gemm.GetCLayout(); }
__device__ static auto GetWaveIdx()
{
const index_t thread_id = get_thread_local_1d_id();
__device__ constexpr auto GetNumBlks() const { return xdlops_gemm.GetCLayout().GetNumBlks(); }
const auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
__device__ constexpr auto GetBlkSize() const { return xdlops_gemm.GetCLayout().GetBlkSize(); }
return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
}
__device__ static auto CalculateAThreadOriginDataIndex()
{
const index_t thread_id = get_thread_local_1d_id();
const index_t waveId = thread_id / WaveSize;
const index_t laneId = thread_id % WaveSize;
const index_t waveId_m = waveId / NWaves;
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto laneId = wave_idx[I2];
const auto blk_idx = xdlops_gemm.GetBlkIdx();
const auto blk_id = blk_idx[I0];
const auto blk_td = blk_idx[I1];
if constexpr(xdlops_gemm.IsKReduction)
{
const index_t m_offset = waveId_m * MPerWave + xdlops_gemm.GetBlkTd(laneId);
const index_t k_offset = xdlops_gemm.GetBlkId(laneId);
return make_tuple(k_offset, 0, m_offset, 0);
return make_tuple(blk_id, 0, waveId_m, blk_td, 0);
}
else
{
const index_t m_offset = waveId_m * MPerWave + laneId;
const index_t k_offset = 0;
return make_tuple(k_offset, 0, m_offset, 0);
return make_tuple(0, 0, waveId_m, laneId, 0);
}
}
__device__ static auto CalculateBThreadOriginDataIndex()
{
const index_t thread_id = get_thread_local_1d_id();
const index_t waveId = thread_id / WaveSize;
const index_t laneId = thread_id % WaveSize;
const index_t waveId_n = waveId % NWaves;
const auto wave_idx = GetWaveIdx();
const auto waveId_n = wave_idx[I1];
const auto laneId = wave_idx[I2];
const auto blk_idx = xdlops_gemm.GetBlkIdx();
const auto blk_id = blk_idx[I0];
const auto blk_td = blk_idx[I1];
if constexpr(xdlops_gemm.IsKReduction)
{
const index_t n_offset = waveId_n * NPerWave + xdlops_gemm.GetBlkTd(laneId);
const index_t k_offset = xdlops_gemm.GetBlkId(laneId);
return make_tuple(k_offset, 0, n_offset, 0);
return make_tuple(blk_id, 0, waveId_n, blk_td, 0);
}
else
{
const index_t n_offset = waveId_n * NPerWave + laneId;
const index_t k_offset = 0;
return make_tuple(k_offset, 0, n_offset, 0);
return make_tuple(0, 0, waveId_n, laneId, 0);
}
}
......@@ -92,263 +109,117 @@ struct BlockwiseGemmXdlops_km_kn_m0m1m2n_v1
__device__ static CIndex
CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto waveId_n = wave_idx[I1];
const index_t waveId = get_thread_local_1d_id() / WaveSize;
const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
const auto thread_mtx_on_blk = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
constexpr auto mrepeat_mwave_mperxdl_to_m = make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{}));
const index_t waveId_m = waveId / NWaves;
const index_t waveId_n = waveId % NWaves;
constexpr auto nrepeat_nwave_nperxdl_to_n = make_naive_tensor_descriptor_packed(
make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerXDL>{}));
const index_t m_offset = m0 * M1 + waveId_m * MPerWave + thread_mtx_on_blk[I0];
const index_t n_offset = n0 * N1 + waveId_n * NPerWave + thread_mtx_on_blk[I1];
const index_t c_thread_m =
mrepeat_mwave_mperxdl_to_m.CalculateOffset(make_tuple(m0, waveId_m, blk_idx[I0]));
const index_t c_thread_n =
nrepeat_nwave_nperxdl_to_n.CalculateOffset(make_tuple(n0, waveId_n, blk_idx[I1]));
return CIndex{m_offset, n_offset};
return CIndex{c_thread_m, c_thread_n};
}
__device__ BlockwiseGemmXdlops_km_kn_m0m1m2n_v1()
: a_thread_copy_{CalculateAThreadOriginDataIndex()},
b_thread_copy_{CalculateBThreadOriginDataIndex()}
__host__ __device__ BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1()
{
static_assert(ABlockDesc::IsKnownAtCompileTime() && BBlockDesc::IsKnownAtCompileTime(),
static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
BK0NK1BlockDesc::IsKnownAtCompileTime(),
"wrong! Desc should be known at compile-time");
static_assert(ABlockDesc{}.GetLength(I0) == BBlockDesc{}.GetLength(I0),
"wrong! K dimension not consistent");
static_assert(AK0MK1BlockDesc{}.GetLength(I0) == BK0NK1BlockDesc{}.GetLength(I0),
"wrong! K0 dimension not consistent");
static_assert(ABlockDesc{}.GetLength(I3) == BBlockDesc{}.GetLength(I3),
static_assert(AK0MK1BlockDesc{}.GetLength(I2) == BK0NK1BlockDesc{}.GetLength(I2),
"wrong! K1 dimension not consistent");
static_assert(BlockSize == MWaves * NWaves * WaveSize,
"BlockSize != MWaves * NWaves * WaveSize\n");
static_assert(K1 == BBlockDesc{}.GetLength(I3), "K1 is wrong!");
constexpr index_t KPerBlock = ABlockDesc{}.GetLength(I0);
static_assert(KPerBlock % xdlops_gemm.KPerXdlops == 0, "KPerBlock is wrong!");
static_assert(K1 % xdlops_gemm.mfma_type.k_base == 0, "K1 is wrong!");
}
template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
__device__ void Run(const ABlockBuffer& a_block_buf,
const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatAB>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatAB>(
b_thread_desc_.GetElementSpaceSize());
constexpr index_t KPerBlock = ABlockDesc{}.GetLength(I0);
vector_type<FloatAB, a_thread_desc_.GetElementSpaceSize()> a_thread_vec;
vector_type<FloatAB, b_thread_desc_.GetElementSpaceSize()> b_thread_vec;
static_for<0, KPerBlock, xdlops_gemm.KPerXdlops>{}([&](auto k) {
// read A
a_thread_copy_.Run(ABlockDesc{},
make_tuple(k, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, I0, I0, I0),
a_thread_buf);
// read B
b_thread_copy_.Run(BBlockDesc{},
make_tuple(k, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, I0, I0, I0),
b_thread_buf);
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.mfma_type.k_base>::type;
static_for<0, a_thread_desc_.GetElementSpaceSize(), 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatAB>()(Number<i>{}) = a_thread_buf[Number<i>{}];
});
static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
"wrong!");
static_for<0, b_thread_desc_.GetElementSpaceSize(), 1>{}([&](auto i) {
b_thread_vec.template AsType<FloatAB>()(Number<i>{}) = b_thread_buf[Number<i>{}];
});
constexpr index_t NumBlks = CXdlopsLayout.GetNumBlks();
constexpr index_t NumXdlops = CXdlopsLayout.GetNumXdlops();
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
xdlops_gemm.template Run<decltype(a_thread_desc_),
decltype(b_thread_desc_),
decltype(c_thread_desc_),
m0,
n0>(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf);
});
});
});
static_assert(NumBlks == 1 && NumXdlops == 1, "K Reduction Mfma only");
}
private:
// A[K, M]
static constexpr auto a_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, Number<MRepeat>{}, I1, Number<K1>{}));
// B[K, N]
static constexpr auto b_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, Number<NRepeat>{}, I1, Number<K1>{}));
static constexpr auto c_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, Number<NRepeat>{}));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
ABlockDesc,
decltype(a_thread_desc_),
Sequence<1, MRepeat, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
K1,
1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
BBlockDesc,
decltype(b_thread_desc_),
Sequence<1, NRepeat, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
K1,
1>;
AThreadCopy a_thread_copy_;
BThreadCopy b_thread_copy_;
};
template <index_t BlockSize,
typename FloatAB,
class ABlockDesc,
class BBlockDesc,
index_t MPerWave,
index_t NPerWave,
index_t K1>
struct BlockwiseGemmXdlops_km_kn_m0m1m2n_v1_2x2pipeline
{
using CIndex = MultiIndex<2>;
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 xdlops_gemm = XdlopsGemm<float, MPerWave, NPerWave, K1>{};
static constexpr index_t WaveSize = 64;
static constexpr index_t M0 = ABlockDesc{}.GetLength(I1);
static constexpr index_t M1 = ABlockDesc{}.GetLength(I2);
static constexpr index_t N0 = BBlockDesc{}.GetLength(I1);
static constexpr index_t N1 = BBlockDesc{}.GetLength(I2);
static constexpr index_t MWaves = M1 / MPerWave;
static constexpr index_t NWaves = N1 / NPerWave;
static constexpr index_t MRepeat = M0;
static constexpr index_t NRepeat = N0;
__device__ constexpr auto GetCLayout() const { return xdlops_gemm.GetCLayout(); }
__device__ constexpr auto GetNumBlks() const { return xdlops_gemm.GetCLayout().GetNumBlks(); }
__device__ constexpr auto GetBlkSize() const { return xdlops_gemm.GetCLayout().GetBlkSize(); }
__device__ static auto CalculateAThreadOriginDataIndex()
{
const index_t thread_id = get_thread_local_1d_id();
const index_t waveId = thread_id / WaveSize;
const index_t laneId = thread_id % WaveSize;
const index_t waveId_m = waveId / NWaves;
if constexpr(xdlops_gemm.IsKReduction)
__host__ __device__ static constexpr auto GetCM0N0M1N1M2M3M4N2BlockDescriptor()
{
const index_t m_offset = waveId_m * MPerWave + xdlops_gemm.GetBlkTd(laneId);
const index_t k_offset = xdlops_gemm.GetBlkId(laneId);
return make_tuple(k_offset, 0, m_offset, 0);
constexpr auto M2 = Number<CXdlopsLayout.M1()>{};
constexpr auto M3 = Number<CXdlopsLayout.N1()>{};
constexpr auto M4 = Number<CXdlopsLayout.M0()>{};
constexpr auto N2 = Number<CXdlopsLayout.N0()>{};
return make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
Number<NRepeat>{},
Number<MWaves>{},
Number<NWaves>{},
Number<M2>{},
Number<M3>{},
Number<M4>{},
Number<N2>{}));
}
else
{
const index_t m_offset = waveId_m * MPerWave + laneId;
const index_t k_offset = 0;
return make_tuple(k_offset, 0, m_offset, 0);
}
}
__device__ static auto CalculateBThreadOriginDataIndex()
{
const index_t thread_id = get_thread_local_1d_id();
const index_t waveId = thread_id / WaveSize;
const index_t laneId = thread_id % WaveSize;
const index_t waveId_n = waveId % NWaves;
if constexpr(xdlops_gemm.IsKReduction)
{
const index_t n_offset = waveId_n * NPerWave + xdlops_gemm.GetBlkTd(laneId);
const index_t k_offset = xdlops_gemm.GetBlkId(laneId);
return make_tuple(k_offset, 0, n_offset, 0);
}
else
template <typename CMNGridDesc>
__host__ __device__ static constexpr auto
MakeCM0N0M1N1M2M3M4N2GridDescriptor(const CMNGridDesc& c_m_n_grid_desc)
{
const index_t n_offset = waveId_n * NPerWave + laneId;
const index_t k_offset = 0;
return make_tuple(k_offset, 0, n_offset, 0);
}
///\To-do: pass CGrid desc transform deep inside xdlops gemm
constexpr auto M2 = Number<CXdlopsLayout.M1()>{};
constexpr auto M3 = Number<CXdlopsLayout.N1()>{};
constexpr auto M4 = Number<CXdlopsLayout.M0()>{};
constexpr auto N2 = Number<CXdlopsLayout.N0()>{};
return transform_tensor_descriptor(
c_m_n_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, M2, M3, M4)),
make_unmerge_transform(make_tuple(NRepeat, NWaves, N2))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4, 5, 6>{}, Sequence<1, 3, 7>{}));
}
template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
__device__ static CIndex
CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
__host__ __device__ static constexpr auto MakeAK0M0M1M2K1BlockDescriptor()
{
const index_t waveId = get_thread_local_1d_id() / WaveSize;
const auto thread_mtx_on_blk = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
const index_t waveId_m = waveId / NWaves;
const index_t waveId_n = waveId % NWaves;
const index_t m_offset = m0 * M1 + waveId_m * MPerWave + thread_mtx_on_blk[I0];
const index_t n_offset = n0 * N1 + waveId_n * NPerWave + thread_mtx_on_blk[I1];
return CIndex{m_offset, n_offset};
return transform_tensor_descriptor(
AK0MK1BlockDesc{},
make_tuple(make_pass_through_transform(Number<KPerBlock>{}),
make_unmerge_transform(
make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{})),
make_pass_through_transform(Number<K1>{})),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}));
}
__device__ BlockwiseGemmXdlops_km_kn_m0m1m2n_v1_2x2pipeline()
: a_thread_copy_{CalculateAThreadOriginDataIndex()},
b_thread_copy_{CalculateBThreadOriginDataIndex()}
__host__ __device__ static constexpr auto MakeBK0N0N1N2K1BlockDescriptor()
{
static_assert(ABlockDesc::IsKnownAtCompileTime() && BBlockDesc::IsKnownAtCompileTime(),
"wrong! Desc should be known at compile-time");
static_assert(ABlockDesc{}.GetLength(I0) == BBlockDesc{}.GetLength(I0),
"wrong! K dimension not consistent");
static_assert(ABlockDesc{}.GetLength(I3) == BBlockDesc{}.GetLength(I3),
"wrong! K1 dimension not consistent");
static_assert(BlockSize == MWaves * NWaves * WaveSize,
"BlockSize != MWaves * NWaves * WaveSize\n");
static_assert(K1 == BBlockDesc{}.GetLength(I3), "K1 is wrong!");
constexpr index_t KPerBlock = ABlockDesc{}.GetLength(I0);
static_assert(KPerBlock % xdlops_gemm.KPerXdlops == 0, "KPerBlock is wrong!");
static_assert(K1 % xdlops_gemm.mfma_type.k_base == 0, "K1 is wrong!");
return transform_tensor_descriptor(
BK0NK1BlockDesc{},
make_tuple(make_pass_through_transform(Number<KPerBlock>{}),
make_unmerge_transform(
make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerXDL>{})),
make_pass_through_transform(Number<K1>{})),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}));
}
static constexpr auto a_k0_m0_m1_m2_k1_block_desc = MakeAK0M0M1M2K1BlockDescriptor();
static constexpr auto b_k0_n0_n1_n2_k1_block_desc = MakeBK0N0N1N2K1BlockDescriptor();
template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
__device__ void Run(const ABlockBuffer& a_block_buf,
const BBlockBuffer& b_block_buf,
......@@ -359,165 +230,88 @@ struct BlockwiseGemmXdlops_km_kn_m0m1m2n_v1_2x2pipeline
auto b_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatAB>(
b_thread_desc_.GetElementSpaceSize());
constexpr index_t KPerBlock = ABlockDesc{}.GetLength(I0);
vector_type<FloatAB, K1> a_thread_vec;
// read A_sub_0
a_thread_copy_.Run(ABlockDesc{},
make_tuple(I0, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, I0, I0, I0),
a_thread_buf);
// read B_sub_0
b_thread_copy_.Run(BBlockDesc{},
make_tuple(I0, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, I0, I0, I0),
b_thread_buf);
// read B_sub_1
b_thread_copy_.Run(BBlockDesc{},
make_tuple(I0, I1, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, I1, I0, I0),
b_thread_buf);
// read A_sub_1
a_thread_copy_.Run(ABlockDesc{},
make_tuple(I0, I1, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, I1, I0, I0),
a_thread_buf);
// C_sub_00 += transpose(A_sub_0) * B_sub_0
xdlops_gemm.template Run<decltype(a_thread_desc_),
decltype(b_thread_desc_),
decltype(c_thread_desc_),
0,
0>(a_thread_buf, b_thread_buf, c_thread_buf);
vector_type<FloatAB, K1> b_thread_vec;
// C_sub_01 += transpose(A_sub_0) * B_sub_1
xdlops_gemm.template Run<decltype(a_thread_desc_),
decltype(b_thread_desc_),
decltype(c_thread_desc_),
0,
1>(a_thread_buf, b_thread_buf, c_thread_buf);
static_for<xdlops_gemm.KPerXdlops, KPerBlock, xdlops_gemm.KPerXdlops>{}([&](auto k) {
// read A_sub_0
a_thread_copy_.Run(ABlockDesc{},
make_tuple(k, I0, I0, I0),
static_for<0, KPerBlock, xdlops_gemm.KPerXdlops>{}([&](auto k) {
// read A
a_thread_copy_.Run(a_k0_m0_m1_m2_k1_block_desc,
make_tuple(k, I0, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, I0, I0, I0),
make_tuple(I0, I0, I0, I0, I0),
a_thread_buf);
// C_sub_10 += transpose(A_sub_1) * B_sub_0
xdlops_gemm.template Run<decltype(a_thread_desc_),
decltype(b_thread_desc_),
decltype(c_thread_desc_),
1,
0>(a_thread_buf, b_thread_buf, c_thread_buf);
// read B_sub_0
b_thread_copy_.Run(BBlockDesc{},
make_tuple(k, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, I0, I0, I0),
b_thread_buf);
// C_sub_11 += transpose(A_sub_1) * B_sub_1
xdlops_gemm.template Run<decltype(a_thread_desc_),
decltype(b_thread_desc_),
decltype(c_thread_desc_),
1,
1>(a_thread_buf, b_thread_buf, c_thread_buf);
// read B_sub_1
b_thread_copy_.Run(BBlockDesc{},
make_tuple(k, I1, I0, I0),
// read B
b_thread_copy_.Run(b_k0_n0_n1_n2_k1_block_desc,
make_tuple(k, I0, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, I1, I0, I0),
make_tuple(I0, I0, I0, I0, I0),
b_thread_buf);
// read A_sub_1
a_thread_copy_.Run(ABlockDesc{},
make_tuple(k, I1, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, I1, I0, I0),
a_thread_buf);
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.mfma_type.k_base>::type;
// C_sub_00 += transpose(A_sub_0) * B_sub_0
xdlops_gemm.template Run<decltype(a_thread_desc_),
decltype(b_thread_desc_),
decltype(c_thread_desc_),
0,
0>(a_thread_buf, b_thread_buf, c_thread_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, K1, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatAB>()(Number<i>{}) = a_thread_buf
[Number<a_thread_desc_.CalculateOffset(make_tuple(0, m0, 0, 0, i))>{}];
});
// C_sub_01 += transpose(A_sub_0) * B_sub_1
xdlops_gemm.template Run<decltype(a_thread_desc_),
decltype(b_thread_desc_),
decltype(c_thread_desc_),
0,
1>(a_thread_buf, b_thread_buf, c_thread_buf);
static_for<0, K1, 1>{}([&](auto i) {
b_thread_vec.template AsType<FloatAB>()(Number<i>{}) = b_thread_buf
[Number<b_thread_desc_.CalculateOffset(make_tuple(0, n0, 0, 0, i))>{}];
});
// C_sub_10 += transpose(A_sub_1) * B_sub_0
xdlops_gemm.template Run<decltype(a_thread_desc_),
decltype(b_thread_desc_),
decltype(c_thread_desc_),
1,
0>(a_thread_buf, b_thread_buf, c_thread_buf);
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// C_sub_11 += transpose(A_sub_1) * B_sub_1
xdlops_gemm.template Run<decltype(a_thread_desc_),
decltype(b_thread_desc_),
decltype(c_thread_desc_),
1,
1>(a_thread_buf, b_thread_buf, c_thread_buf);
xdlops_gemm.template Run<c_offset>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf);
});
});
});
}
private:
// A[K, M]
static constexpr auto a_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, Number<MRepeat>{}, I1, Number<K1>{}));
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(I1, Number<MRepeat>{}, I1, I1, Number<K1>{}));
// B[K, N]
static constexpr auto b_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, Number<NRepeat>{}, I1, Number<K1>{}));
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(I1, Number<NRepeat>{}, I1, I1, Number<K1>{}));
static constexpr auto c_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, Number<NRepeat>{}));
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, Number<xdlops_gemm.GetNumXdlops()>{}));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
ABlockDesc,
decltype(a_k0_m0_m1_m2_k1_block_desc),
decltype(a_thread_desc_),
Sequence<1, 1, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
1, // K1,
Sequence<1, MRepeat, 1, 1, K1>,
Sequence<0, 1, 2, 3, 4>,
4,
K1,
1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
BBlockDesc,
decltype(b_k0_n0_n1_n2_k1_block_desc),
decltype(b_thread_desc_),
Sequence<1, 1, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
1, // K1,
Sequence<1, NRepeat, 1, 1, K1>,
Sequence<0, 1, 2, 3, 4>,
4,
K1,
1>;
AThreadCopy a_thread_copy_;
BThreadCopy b_thread_copy_;
AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
};
} // namespace ck
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r3.hpp
View file @ d3341a67
......@@ -18,7 +18,7 @@ template <typename GridwiseGemm,
typename FloatC,
typename AK0MK1GridDesc,
typename BK0NK1GridDesc,
typename CM0M1M2NGridDesc,
typename CM0N0M1N1M2M3M4N2GridDesc,
typename CBlockClusterAdaptor>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
......@@ -29,7 +29,7 @@ __global__ void
FloatC* __restrict__ p_c_grid,
const AK0MK1GridDesc a_k0_m_k1_grid_desc,
const BK0NK1GridDesc b_k0_n_k1_grid_desc,
const CM0M1M2NGridDesc c_m0_m1_m2_n_grid_desc,
const CM0N0M1N1M2M3M4N2GridDesc c_m0_m1_m2_n_grid_desc,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{
constexpr index_t shared_block_size =
......@@ -43,7 +43,7 @@ __global__ void
p_shared_block,
a_k0_m_k1_grid_desc,
b_k0_n_k1_grid_desc,
c_m0_m1_m2_n_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_block_cluster_adaptor);
}
#elif CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER
......@@ -52,7 +52,7 @@ template <typename GridwiseGemm,
typename FloatC,
typename AK0MK1GridDesc,
typename BK0NK1GridDesc,
typename CM0M1M2NGridDesc,
typename CM0N0M1N1M2M3M4N2GridDesc,
typename CBlockClusterAdaptor>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
......@@ -63,7 +63,7 @@ __global__ void
FloatC* __restrict__ p_c_grid,
const void CONSTANT* p_a_k0_m_k1_grid_desc,
const void CONSTANT* p_b_k0_n_k1_grid_desc,
const void CONSTANT* p_c_m0_m1_m2_n_grid_desc,
const void CONSTANT* p_c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
const void CONSTANT* p_c_block_cluster_adaptor)
{
constexpr index_t shared_block_size =
......@@ -73,8 +73,9 @@ __global__ void
cast_pointer_to_generic_address_space(p_a_k0_m_k1_grid_desc));
const auto b_k0_n_k1_grid_desc = *reinterpret_cast<const BK0NK1GridDesc*>(
cast_pointer_to_generic_address_space(p_b_k0_n_k1_grid_desc));
const auto c_m0_m1_m2_n_grid_desc = *reinterpret_cast<const CM0M1M2NGridDesc*>(
cast_pointer_to_generic_address_space(p_c_m0_m1_m2_n_grid_desc));
const auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc =
*reinterpret_cast<const CM0N0M1N1M2M3M4N2GridDesc*>(
cast_pointer_to_generic_address_space(p_c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc));
const auto c_block_cluster_adaptor = *reinterpret_cast<const CBlockClusterAdaptor*>(
cast_pointer_to_generic_address_space(p_c_block_cluster_adaptor));
......@@ -86,7 +87,7 @@ __global__ void
p_shared_block,
a_k0_m_k1_grid_desc,
b_k0_n_k1_grid_desc,
c_m0_m1_m2_n_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_block_cluster_adaptor);
}
#endif
......@@ -138,6 +139,9 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
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>{};
// K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{};
......@@ -201,29 +205,28 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
}
__host__ __device__ static constexpr auto
MakeCM0M1M2NGridDescriptor(const CMNGridDesc& c_m_n_grid_desc)
MakeCM0N0M1N1M2M3M4N2GridDescriptor(const CMNGridDesc& c_m_n_grid_desc)
{
constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerWave, NPerWave, K1>{};
constexpr auto CLayout = xdlops_gemm.GetCLayout();
constexpr auto M0 = Number<CLayout.M1()>{};
constexpr auto M1 = Number<CLayout.N1()>{};
constexpr auto M2 = Number<CLayout.M0()>{};
constexpr auto max_lds_align = K1;
constexpr index_t MWaves = MPerBlock / (MPerWave * MRepeat);
constexpr index_t NWaves = NPerBlock / (NPerWave * NRepeat);
constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
constexpr auto N1 = Number<CLayout.N0()>{};
constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
const auto c_m0_m1_m2_n_grid_desc = transform_tensor_descriptor(
c_m_n_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, M0, M1, M2)),
make_unmerge_transform(make_tuple(NRepeat, NWaves, N1))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4, 5, 6>{}, Sequence<1, 3, 7>{}));
using BlockwiseGemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
decltype(a_k0_m_k1_block_desc),
decltype(b_k0_n_k1_block_desc),
MPerWave,
NPerWave,
MRepeat,
NRepeat,
K1>;
return c_m0_m1_m2_n_grid_desc;
return BlockwiseGemm::MakeCM0N0M1N1M2M3M4N2GridDescriptor(c_m_n_grid_desc);
}
__host__ __device__ static constexpr auto
......@@ -253,7 +256,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
return c_blockid_to_m0_n0_block_cluster_adaptor;
}
using CM0M1M2NGridDesc = decltype(MakeCM0M1M2NGridDescriptor(CMNGridDesc{}));
using CM0N0M1N1M2M3M4N2GridDesc = decltype(MakeCM0N0M1N1M2M3M4N2GridDescriptor(CMNGridDesc{}));
using CBlockClusterAdaptor = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}));
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
......@@ -262,7 +265,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
FloatAB* __restrict__ p_shared_block,
const AK0MK1GridDesc& a_k0_m_k1_grid_desc,
const BK0NK1GridDesc& b_k0_n_k1_grid_desc,
const CM0M1M2NGridDesc& c_m0_m1_m2_n_grid_desc,
const CM0N0M1N1M2M3M4N2GridDesc& c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
const CBlockClusterAdaptor& c_block_cluster_adaptor)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
......@@ -270,7 +273,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_b_grid, b_k0_n_k1_grid_desc.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_c_grid, c_m0_m1_m2_n_grid_desc.GetElementSpaceSize());
p_c_grid, c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetElementSpaceSize());
const auto K0 = a_k0_m_k1_grid_desc.GetLength(I0);
......@@ -358,50 +361,26 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
// register
// sanity check
static_assert(MPerBlock % (MPerWave * MRepeat) == 0 &&
NPerBlock % (NPerWave * NRepeat) == 0,
"wrong!");
constexpr auto a_k0_m0_m1_k1_block_desc = transform_tensor_descriptor(
a_k0_m_k1_block_desc,
make_tuple(make_pass_through_transform(Number<KPerBlock>{}),
make_unmerge_transform(
make_tuple(Number<MRepeat>{}, Number<MPerBlock / MRepeat>{})),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
constexpr auto b_k0_n0_n1_k1_block_desc = transform_tensor_descriptor(
b_k0_n_k1_block_desc,
make_tuple(make_pass_through_transform(Number<KPerBlock>{}),
make_unmerge_transform(
make_tuple(Number<NRepeat>{}, Number<NPerBlock / NRepeat>{})),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
const auto blockwise_gemm =
BlockwiseGemmXdlops_km_kn_m0m1m2n_v1<BlockSize,
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
decltype(a_k0_m0_m1_k1_block_desc),
decltype(b_k0_n0_n1_k1_block_desc),
decltype(a_k0_m_k1_block_desc),
decltype(b_k0_n_k1_block_desc),
MPerWave,
NPerWave,
MRepeat,
NRepeat,
K1>{};
constexpr auto CLayout = blockwise_gemm.GetCLayout();
constexpr index_t BlkSize = CLayout.GetBlkSize();
constexpr index_t NumBlks = CLayout.GetNumBlks();
constexpr index_t NumXdlops = CLayout.GetNumXdlops();
static_assert(NumBlks == 1 && NumXdlops == 1, "K Reduction Mfma only");
constexpr auto c_mr_nr_blk_desc =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, Number<NRepeat>{}));
constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
blockwise_gemm.GetCM0N0M1N1M2M3M4N2ThreadDesc();
constexpr auto CBlkSize = c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc.GetElementSpaceSize();
StaticBuffer<AddressSpaceEnum_t::Vgpr,
vector_type<FloatAcc, BlkSize>,
vector_type<FloatAcc, CBlkSize>,
c_mr_nr_blk_desc.GetElementSpaceSize(),
true>
c_thread_buf;
......@@ -474,94 +453,14 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
}
#if 0
// output: register to global memory
{
constexpr index_t M0 = CLayout.M1();
constexpr index_t M1 = CLayout.N1();
constexpr index_t M2 = CLayout.M0();
constexpr index_t N0 = CLayout.N1();
constexpr index_t N1 = CLayout.N0();
constexpr auto c_m0_m1_m2_n_thread_desc =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
Number<NRepeat>{},
Number<1>{},
Number<1>{},
Number<M0>{},
Number<1>{},
Number<M2>{},
Number<1>{}));
StaticBuffer<AddressSpaceEnum_t::Vgpr, FloatC, c_m0_m1_m2_n_thread_desc.GetElementSpaceSize(), true>
c_blk_buf_;
static_for<0, MRepeat, 1>{}([&](auto mr_i) {
static_for<0, NRepeat, 1>{}([&](auto nr_i) {
constexpr auto blk_off =
c_mr_nr_blk_desc.CalculateOffset(make_tuple(mr_i, nr_i));
static_for<0, BlkSize, 1>{}([&](auto j) {
c_blk_buf_(Number<blk_off * BlkSize + j>{}) =
c_thread_buf[Number<blk_off>{}]
.template AsType<FloatAcc>()[Number<j>{}];
});
});
});
constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
blockwise_gemm.GetCM0N0M1N1M2M3M4N2BlockDescriptor();
// 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, I0, I0);
const index_t m_thread_data_on_grid =
m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_grid =
n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
constexpr auto c_m0_m1_m2_n_grid_tensor_step_hacks = CGridStepHacks{};
constexpr index_t MWaves = MPerBlock / (MPerWave * MRepeat);
constexpr index_t NWaves = NPerBlock / (NPerWave * NRepeat);
ThreadwiseTensorSliceTransfer_v1r3<
FloatC,
FloatC,
decltype(c_m0_m1_m2_n_thread_desc),
decltype(c_m0_m1_m2_n_grid_desc),
Sequence<MRepeat, NRepeat, 1, 1, M0, 1, M2, 1>,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
CGlobalMemoryDataOperation,
1,
true>{
c_m0_m1_m2_n_grid_desc,
make_multi_index(m_thread_data_on_grid / (M2 * M1 * M0 * MWaves),
n_thread_data_on_grid / (N1 * NWaves),
m_thread_data_on_grid % (M2 * M1 * M0 * MWaves) / (M2 * M1 * M0),
n_thread_data_on_grid % (N1 * NWaves) / N1,
m_thread_data_on_grid % (M2 * M1 * M0) / (M2 * M1),
m_thread_data_on_grid % (M2 * M1) / M2,
m_thread_data_on_grid % M2,
n_thread_data_on_grid % N1)}
.Run(c_m0_m1_m2_n_thread_desc,
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
c_blk_buf_,
c_m0_m1_m2_n_grid_desc,
c_grid_buf,
c_m0_m1_m2_n_grid_tensor_step_hacks);
}
#else
{
constexpr index_t M0 = CLayout.M1();
constexpr index_t M1 = CLayout.N1();
constexpr index_t M2 = CLayout.M0();
constexpr auto c_m0_m1_m2_n_thread_desc = make_naive_tensor_descriptor_packed(
make_tuple(I1, I1, I1, I1, Number<M0>{}, Number<1>{}, Number<M2>{}, Number<1>{}));
constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
......@@ -574,92 +473,96 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
const index_t n_thread_data_on_grid =
n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
constexpr auto c_m0_m1_m2_n_grid_tensor_step_hacks = CGridStepHacks{};
constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_iterator_hacks = CGridStepHacks{};
auto c_thread_copy =
ThreadwiseTensorSliceTransfer_v1r3<FloatC,
FloatC,
decltype(c_m0_m1_m2_n_thread_desc),
decltype(c_m0_m1_m2_n_grid_desc),
Sequence<1, 1, 1, 1, M0, 1, M2, 1>,
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc),
Sequence<I1, I1, I1, I1, M2, I1, M4, I1>,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
CGlobalMemoryDataOperation,
1,
true>{
c_m0_m1_m2_n_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
make_multi_index(0,
0,
0,
0,
m_thread_data_on_grid / (M2 * M1),
m_thread_data_on_grid % (M2 * M1) / M2,
m_thread_data_on_grid % M2,
m_thread_data_on_grid / (M3 * M4),
m_thread_data_on_grid % (M3 * M4) / M4,
m_thread_data_on_grid % M4,
n_thread_data_on_grid)};
auto init_copy = [&](auto c_thread_idx_) {
constexpr auto blk_off = c_mr_nr_blk_desc.CalculateOffset(c_thread_idx_);
c_thread_copy.Run(c_m0_m1_m2_n_thread_desc,
c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
c_thread_buf[Number<blk_off>{}].template AsType<FloatAcc>(),
c_m0_m1_m2_n_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_grid_buf,
c_m0_m1_m2_n_grid_tensor_step_hacks);
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_iterator_hacks);
return c_thread_idx_;
};
auto mrepeat_plus_copy = [&](auto c_thread_idx_) {
constexpr auto mrepeat_step_plus = make_multi_index(1, 0, 0, 0, 0, 0, 0, 0);
c_thread_copy.MoveDstSliceWindow(c_m0_m1_m2_n_grid_desc, mrepeat_step_plus);
c_thread_copy.MoveDstSliceWindow(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
mrepeat_step_plus);
constexpr auto blk_off = c_mr_nr_blk_desc.CalculateOffset(c_thread_idx_);
c_thread_copy.Run(c_m0_m1_m2_n_thread_desc,
c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
c_thread_buf[Number<blk_off>{}].template AsType<FloatAcc>(),
c_m0_m1_m2_n_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_grid_buf,
c_m0_m1_m2_n_grid_tensor_step_hacks);
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_iterator_hacks);
};
auto nrepeat_plus_copy = [&](auto c_thread_idx_) {
constexpr auto nrepeat_step_plus = make_multi_index(0, 1, 0, 0, 0, 0, 0, 0);
c_thread_copy.MoveDstSliceWindow(c_m0_m1_m2_n_grid_desc, nrepeat_step_plus);
c_thread_copy.MoveDstSliceWindow(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
nrepeat_step_plus);
constexpr auto blk_off = c_mr_nr_blk_desc.CalculateOffset(c_thread_idx_);
c_thread_copy.Run(c_m0_m1_m2_n_thread_desc,
c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
c_thread_buf[Number<blk_off>{}].template AsType<FloatAcc>(),
c_m0_m1_m2_n_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_grid_buf,
c_m0_m1_m2_n_grid_tensor_step_hacks);
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_iterator_hacks);
};
auto mrepeat_minus_copy = [&](auto c_thread_idx_) {
constexpr auto mrepeat_step_plus = make_multi_index(-1, 0, 0, 0, 0, 0, 0, 0);
c_thread_copy.MoveDstSliceWindow(c_m0_m1_m2_n_grid_desc, mrepeat_step_plus);
c_thread_copy.MoveDstSliceWindow(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
mrepeat_step_plus);
constexpr auto blk_off = c_mr_nr_blk_desc.CalculateOffset(c_thread_idx_);
c_thread_copy.Run(c_m0_m1_m2_n_thread_desc,
c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
c_thread_buf[Number<blk_off>{}].template AsType<FloatAcc>(),
c_m0_m1_m2_n_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_grid_buf,
c_m0_m1_m2_n_grid_tensor_step_hacks);
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_iterator_hacks);
};
auto nrepeat_minus_copy = [&](auto c_thread_idx_) {
constexpr auto nrepeat_step_minus = make_multi_index(0, -1, 0, 0, 0, 0, 0, 0);
c_thread_copy.MoveDstSliceWindow(c_m0_m1_m2_n_grid_desc, nrepeat_step_minus);
c_thread_copy.MoveDstSliceWindow(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
nrepeat_step_minus);
constexpr auto blk_off = c_mr_nr_blk_desc.CalculateOffset(c_thread_idx_);
c_thread_copy.Run(c_m0_m1_m2_n_thread_desc,
c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
c_thread_buf[Number<blk_off>{}].template AsType<FloatAcc>(),
c_m0_m1_m2_n_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_grid_buf,
c_m0_m1_m2_n_grid_tensor_step_hacks);
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_iterator_hacks);
};
static_assert((MRepeat == 4 && NRepeat == 4) or (MRepeat == 4 && NRepeat == 2) or
......@@ -791,7 +694,6 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
init_copy(make_tuple(I0, I0));
}
}
#endif
}
}; // namespace ck
......
composable_kernel/include/tensor_operation/xdlops_gemm.hpp
View file @ d3341a67
......@@ -709,19 +709,59 @@ struct XdlopsGemm
static_assert(mfma_type.k % mfma_type.k_base == 0, "k % kbase != 0!");
}
template <typename CM0N0M1N1M2N2GridDesc>
__host__ __device__ static constexpr auto
MakeCM0N0M1N1M2M3M4N2GridDescriptor(const CM0N0M1N1M2N2GridDesc& c_m0_n0_m1_n1_m2_n2_grid_desc)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto I4 = Number<4>{};
constexpr auto I5 = Number<5>{};
constexpr auto M0 = c_m0_n0_m1_n1_m2_n2_grid_desc.GetLength(I0);
constexpr auto N0 = c_m0_n0_m1_n1_m2_n2_grid_desc.GetLength(I1);
constexpr auto M1 = c_m0_n0_m1_n1_m2_n2_grid_desc.GetLength(I2);
constexpr auto N1 = c_m0_n0_m1_n1_m2_n2_grid_desc.GetLength(I3);
constexpr auto M2 = c_m0_n0_m1_n1_m2_n2_grid_desc.GetLength(I4);
constexpr auto N2 = c_m0_n0_m1_n1_m2_n2_grid_desc.GetLength(I5);
static_assert(N2 == mfma_type.num_threads_blk, "");
static_assert(
M2 == (mfma_type.num_groups_blk * mfma_type.num_output_blks * mfma_type.group_size),
"");
return transform_dynamic_tensor_descriptor(
c_m0_n0_m1_n1_m2_n2_grid_desc,
make_tuple(make_pass_through_transform(M0),
make_pass_through_transform(N0),
make_pass_through_transform(M1),
make_pass_through_transform(N1),
make_unmerge_transform(make_tuple(mfma_type.num_groups_blk,
mfma_type.num_input_blks,
mfma_type.group_size)),
make_pass_through_transform(mfma_type.num_threads_blk)),
make_tuple(Sequence<0>{},
Sequence<1>{},
Sequence<2>{},
Sequence<3>{},
Sequence<4>{},
Sequence<5>{}),
make_tuple(Sequence<0>{},
Sequence<1>{},
Sequence<2>{},
Sequence<3>{},
Sequence<4, 5, 6>{},
Sequence<7>{}));
}
__device__ static constexpr index_t GetRegSizePerXdlops()
{
return MPerXdlops * NPerXdlops / mfma_type.wave_size;
}
template <class ADesc,
class BDesc,
class CDesc,
index_t m0,
index_t n0,
class FloatA,
class FloatB,
class FloatC>
template <index_t c_offset, class FloatA, class FloatB, class FloatC>
__device__ void Run(const FloatA& p_a_wave, const FloatB& p_b_wave, FloatC& p_c_thread) const
{
static_assert(is_same<base_type, float>::value || is_same<base_type, half_t>::value ||
......@@ -730,24 +770,35 @@ struct XdlopsGemm
static_assert(KPack % mfma_type.k_base == 0, "KPack cannot be divided by k_base");
constexpr index_t c_offset = CDesc{}.CalculateOffset(make_tuple(m0, n0)) * GetNumXdlops();
static_for<0, KPack, mfma_type.k_base>{}([&](auto k) {
constexpr index_t a_offset = ADesc{}.CalculateOffset(make_tuple(0, m0, 0, k));
constexpr index_t b_offset = BDesc{}.CalculateOffset(make_tuple(0, n0, 0, k));
static_for<0, KPack / mfma_type.k_base, 1>{}([&](auto k) {
mfma_type.template run<MPerXdlops, NPerXdlops, c_offset>(
p_a_wave[Number<a_offset / mfma_type.k_base>{}],
p_b_wave[Number<b_offset / mfma_type.k_base>{}],
p_c_thread);
p_a_wave[k], p_b_wave[k], p_c_thread);
});
}
static constexpr auto GetBlkIdx()
{
const auto threadidx_to_blk_idx_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(
make_tuple(1, mfma_type.num_input_blks, mfma_type.num_threads_blk))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto blk_idx = threadidx_to_blk_idx_adaptor.CalculateBottomIndex(
make_multi_index(get_thread_local_1d_id()));
const auto blk_id = blk_idx[Number<1>{}];
const auto blk_td = blk_idx[Number<2>{}];
return make_tuple(blk_id, blk_td);
}
__device__ static CIndex GetBeginOfThreadBlk(index_t xdlops_i, index_t blk_i)
{
const index_t laneId = get_thread_local_1d_id() % mfma_type.wave_size;
const index_t blk_id = laneId / mfma_type.num_threads_blk;
const index_t blk_td = laneId % mfma_type.num_threads_blk;
const auto blk_idx = GetBlkIdx();
const auto blk_id = blk_idx[Number<0>{}];
const auto blk_td = blk_idx[Number<1>{}];
index_t n_offset = blk_i * mfma_type.n + blk_td;
index_t m_offset = xdlops_i * mfma_type.m + blk_id * mfma_type.group_size;
......@@ -755,24 +806,12 @@ struct XdlopsGemm
return CIndex{m_offset, n_offset};
}
static constexpr index_t MRepeats = GetXdlopsInfo().MRepeats;
static constexpr index_t NRepeats = GetXdlopsInfo().NRepeats;
static constexpr index_t MPerXdlops = GetXdlopsInfo().MPerXdlops;
static constexpr index_t NPerXdlops = GetXdlopsInfo().NPerXdlops;
static constexpr index_t KPerXdlops = GetXdlopsInfo().GetKPerXdlops();
static constexpr bool IsKReduction = GetXdlopsInfo().IsKReduction();
static constexpr bool IsABroadcast = GetXdlopsInfo().IsABroadcast();
static constexpr index_t KPerXdlops = GetXdlopsInfo().GetKPerXdlops();
static constexpr auto GetBlkId(const index_t lane_id)
{
return lane_id / mfma_type.num_threads_blk;
}
static constexpr auto GetBlkTd(const index_t lane_id)
{
return lane_id % mfma_type.num_threads_blk;
}
static constexpr auto mfma_type = GetXdlopsInfo().mfma_type;
......@@ -794,7 +833,7 @@ struct XdlopsGemm
}
};
__host__ __device__ static constexpr auto GetCLayout() { return CLayout{}; }
__host__ __device__ static constexpr auto GetCXdlopsLayout() { return CLayout{}; }
};
} // namespace ck
......
host/driver_offline/include/driver_gemm_xdlops_v2r3.hpp
View file @ d3341a67
......@@ -129,9 +129,10 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid,
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting");
}
const auto c_m0_m1_m2_n_grid_desc = GridwiseGemm::MakeCM0M1M2NGridDescriptor(c_m_n_grid_desc);
const auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc =
GridwiseGemm::MakeCM0N0M1N1M2M3M4N2GridDescriptor(c_m_n_grid_desc);
using CM0M1M2NGridDesc = decltype(c_m0_m1_m2_n_grid_desc);
using CM0N0M1N1M2M3M4N2GridDesc = decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc);
const auto c_block_cluster_adaptor = GridwiseGemm::MakeCBlockClusterAdaptor(c_m_n_grid_desc);
......@@ -144,7 +145,7 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid,
FloatC,
remove_reference_t<AK0MK1GridDesc>,
remove_reference_t<BK0NK1GridDesc>,
remove_reference_t<CM0M1M2NGridDesc>,
remove_reference_t<CM0N0M1N1M2M3M4N2GridDesc>,
remove_reference_t<CBlockClusterAdaptor>>;
#if CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VALUE
......@@ -158,18 +159,18 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid,
p_c_grid,
a_k0_m_k1_grid_desc,
b_k0_n_k1_grid_desc,
c_m0_m1_m2_n_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_block_cluster_adaptor);
#elif CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER
DeviceMem a_k0_m_k1_grid_desc_dev_buf(sizeof(AK0MK1GridDesc));
DeviceMem b_k0_n_k1_grid_desc_dev_buf(sizeof(BK0NK1GridDesc));
DeviceMem c_m0_m1_m2_n_grid_desc_dev_buf(sizeof(CM0M1M2NGridDesc));
DeviceMem c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc_dev_buf(sizeof(CM0N0M1N1M2M3M4N2GridDesc));
DeviceMem c_block_cluster_adaptor_dev_buf(sizeof(CBlockClusterAdaptor));
a_k0_m_k1_grid_desc_dev_buf.ToDevice(&a_k0_m_k1_grid_desc);
b_k0_n_k1_grid_desc_dev_buf.ToDevice(&b_k0_n_k1_grid_desc);
c_m0_m1_m2_n_grid_desc_dev_buf.ToDevice(&c_m0_m1_m2_n_grid_desc);
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc_dev_buf.ToDevice(&c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc);
c_block_cluster_adaptor_dev_buf.ToDevice(&c_block_cluster_adaptor);
float ave_time = launch_and_time_kernel(
......@@ -183,7 +184,8 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid,
p_c_grid,
cast_pointer_to_constant_address_space(a_k0_m_k1_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(b_k0_n_k1_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(c_m0_m1_m2_n_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(c_block_cluster_adaptor_dev_buf.GetDeviceBuffer()));
#endif
return ave_time;
......
host/driver_offline/src/conv_fwd_driver_offline.cpp
View file @ d3341a67
......@@ -24,7 +24,7 @@
#define USE_CONV_FWD_V4R4R2_NHWC 1
#define USE_CONV_FWD_V6R1_NCHW 0
#define USE_CONV_FWD_V5R1_NCHW 0
#define USE_CONV_FWD_V4R4R2_XDL_NCHW 0
#define USE_CONV_FWD_V4R4R2_XDL_NCHW 1
#define USE_CONV_FWD_V4R4R4_XDL_NHWC 0
enum ConvForwardAlgo
......
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