Skip to content

GitLab

Commit cb46ef7a authored by Po-Yen, Chen's avatar Po-Yen, Chen
Browse files

Remove no-longer used descriptor-creation methods

parent 0cf90eaf
Hide whitespace changes
Inline Side-by-side
Showing with 19 additions and 263 deletions
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
View file @ cb46ef7a
...@@ -82,221 +82,6 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout, ...@@ -82,221 +82,6 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
static constexpr auto I1 = Number<1>{}; static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{}; static constexpr auto I2 = Number<2>{};
static auto MakeAGridDescriptor_AK0_M_AK1(
index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
{
const auto a_grid_desc_mraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both M and K
const auto a_grid_desc_m_k =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(MPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad M, but not K
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_right_pad_transform(M, MPad - M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad K, but not M
const auto a_grid_desc_m_k = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
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>{}));
return a_grid_desc_ak0_m_ak1;
}
else
{
// not pad M or K
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
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>{}));
return a_grid_desc_ak0_m_ak1;
}
}
static auto MakeBGridDescriptor_BK0_N_BK1(
index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
{
const auto b_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
}
}();
if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both N and K
const auto b_grid_desc_n_k =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_right_pad_transform(N, NPad - N),
make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(NPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad N, but not K
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad K, but not N
const auto b_grid_desc_n_k = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
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>{}));
return b_grid_desc_bk0_n_bk1;
}
else
{
// not pad N or K
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
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>{}));
return b_grid_desc_bk0_n_bk1;
}
}
static auto
MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
{
const auto c_grid_desc_mraw_nraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad M, but not N
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad N, but not M
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
// not pad M or N
return c_grid_desc_mraw_nraw;
}
}
using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1, 1, 1, 1));
using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1, 1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1, 1, 1));
// GridwiseGemm // GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1< using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
ALayout, ALayout,
...@@ -345,6 +130,12 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout, ...@@ -345,6 +130,12 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
LoopSched, LoopSched,
PipelineVer>; PipelineVer>;
using AGridDesc_AK0_M_AK1 =
decltype(GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(1, 1, 1, 1, 1, 1));
using BGridDesc_BK0_N_BK1 =
decltype(GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(1, 1, 1, 1, 1, 1));
using CGridDesc_M_N = decltype(GridwiseGemm::MakeCGridDescriptor_M_N(1, 1, 1, 1, 1));
// Argument // Argument
struct Argument : public BaseArgument struct Argument : public BaseArgument
{ {
...@@ -396,56 +187,21 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout, ...@@ -396,56 +187,21 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
__host__ __device__ void Print() const __host__ __device__ void Print() const
{ {
printf("arg {M: %d, N: %d, K: %d, " std::cout << "arg {"
"SA: %d, SB: %d, SC: %d, " << "M:" << M << ", "
"MP: %d, NP: %d, KP: %d, " << "N:" << N << ", "
"AK0: %d, BK0: %d}\n", << "K:" << K << ", "
M, << "SA:" << StrideA << ", "
N, << "SB:" << StrideB << ", "
K, << "SC:" << StrideC << ", "
StrideA, << "MP:" << MPadded << ", "
StrideB, << "NP:" << NPadded << ", "
StrideC, << "KP:" << KPadded << ", "
MPadded, << "AK0:" << AK0 << ", "
NPadded, << "BK0:" << BK0 << "}" << std::endl;
KPadded,
AK0,
BK0);
// std::cout << "arg {"
// << "M:" << M << ", "
// << "N:" << N << ", "
// << "K:" << K << ", "
// << "SA:" << StrideA << ", "
// << "SB:" << StrideB << ", "
// << "SC:" << StrideC << ", "
// << "MP:" << MPadded << ", "
// << "NP:" << NPadded << ", "
// << "KP:" << KPadded << ", "
// << "AK0:" << AK0 << ", "
// << "BK0:" << BK0 << "}" << std::endl;
} }
__host__ __device__ Argument(const Argument& other) __host__ __device__ Argument(const Argument&) = default;
: p_a_grid{other.p_a_grid},
p_b_grid{other.p_b_grid},
p_c_grid{other.p_c_grid},
M{other.M},
N{other.N},
K{other.K},
StrideA{other.StrideA},
StrideB{other.StrideB},
StrideC{other.StrideC},
MPadded{other.MPadded},
NPadded{other.NPadded},
KPadded{other.KPadded},
AK0{other.AK0},
BK0{other.BK0},
a_grid_desc_ak0_m_ak1{other.a_grid_desc_ak0_m_ak1},
b_grid_desc_bk0_n_bk1{other.b_grid_desc_bk0_n_bk1},
c_grid_desc_m_n{other.c_grid_desc_m_n}
{
}
__host__ __device__ ~Argument() override {} __host__ __device__ ~Argument() override {}
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment