Skip to content

GitLab

Commit 0eb75e21 authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into ck_tile/moe

parents 1b4b640b c8b6b642
Expand all Show whitespace changes
Inline Side-by-side
Showing with 2330 additions and 421 deletions
example/ck_tile/01_fmha/script/smoke_test_bwd.sh
View file @ 0eb75e21
......@@ -11,18 +11,19 @@ COMMON_ARGS='-v=1'
set -x
for prec in "fp16" "bf16" ; do
for perm in 0 1 ; do
for hdim in 32 64 128 ; do
for hdim in 32 64 128 256 ; do
for mode in 0 1 ; do
for bias in "n" "e" "a"; do
for dbias in 0 1 ; do
for p_drop in 0.0 0.2; do
for bias in "n" "a" ; do
for dbias in 0 ; do
for p_drop in 0.0 0.2 ; do
for deterministic in 0 ; do
$EXE -prec=$prec -b=1 -h=4 -h_k=2 -d=$hdim -s=259 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=2 -d=$hdim -s=516 -s_k=253 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=4 -h_k=1 -d=$hdim -s=500 -s_k=251 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=1 -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=2 -d=$hdim -s=900 -s_k=258 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=2 -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=1 -d=$hdim -s=987 -s_k=219 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=t:128,30 -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=3 -h_k=1 -d=$hdim -s=244 -s_k=499 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=b:4,35 -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=4 -h_k=2 -d=$hdim -s=259 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=2 -d=$hdim -s=516 -s_k=253 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=4 -h_k=1 -d=$hdim -s=500 -s_k=251 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=1 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=2 -d=$hdim -s=900 -s_k=258 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=2 -v=1 -deterministic=$deterministic -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=1 -d=$hdim -s=987 -s_k=219 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=t:128,30 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=3 -h_k=1 -d=$hdim -s=244 -s_k=499 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=b:4,35 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
done
done
......@@ -31,4 +32,5 @@ done
done
done
done
done
set +x
include/ck/host_utility/device_prop.hpp
View file @ 0eb75e21
......@@ -65,6 +65,12 @@ inline bool is_lds_direct_load_supported()
ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942";
}
inline bool is_bf16_atomic_supported()
{
return ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
ck::get_device_name() == "gfx942";
}
inline bool is_gfx101_supported()
{
return ck::get_device_name() == "gfx1010" || ck::get_device_name() == "gfx1011" ||
......
include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
View file @ 0eb75e21
......@@ -53,6 +53,49 @@ struct DeviceGemmMultipleD : public BaseOperator
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
// GEMM:
// input : A[M, K], B[K, N],
// input : D0[M, N], D1[M, N], ...
// output : E[M, N]
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
// Assume:
// D0, D1, ... and E have the same layout
template <typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceGemmMultipleDSplitK : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t StrideA,
ck::index_t StrideB,
std::array<ck::index_t, NumDTensor> StrideDs,
ck::index_t StrideE,
ck::index_t KBatch,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp
View file @ 0eb75e21
......@@ -126,6 +126,29 @@ struct DeviceGroupedConvFwdMultipleABD : public BaseOperator
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) = 0;
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(APointers p_a,
BPointers p_b,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
const std::array<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_lengths,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_dilations,
const std::array<long_index_t, NDimSpatial>& input_left_pads,
const std::array<long_index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
......
include/ck/tensor_operation/gpu/device/impl/codegen_device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp
View file @ 0eb75e21
......@@ -359,14 +359,14 @@ struct CodegenDeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
using GemmToConvFwdTransformer = TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>;
using ConvToGemmFwdTransformer = TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>;
static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
template <typename ALay>
__host__ __device__ static auto
MakeAGridDescriptor_M_K(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
MakeAGridDescriptor_M_K(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto in_gemmmraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>();
......@@ -379,7 +379,7 @@ struct CodegenDeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
template <typename BLay>
__host__ __device__ static auto
MakeBGridDescriptor_N_K(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
MakeBGridDescriptor_N_K(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto wei_gemmnraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>();
......@@ -392,7 +392,7 @@ struct CodegenDeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
template <typename ELay>
__host__ __device__ static auto
MakeEGridDescriptor_M_N(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
MakeEGridDescriptor_M_N(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto out_gemmmraw_gemmnraw_desc =
conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>();
......@@ -405,7 +405,7 @@ struct CodegenDeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
// Shape of Ds and E must be aligned. Strides can be different.
// Pass e_g_n_k_wos_lengths for logical broadcast.
static auto MakeDsGridDescriptor_M_N(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeDsGridDescriptor_M_N(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
return generate_tuple(
[&](auto i) {
......@@ -417,7 +417,7 @@ struct CodegenDeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
}
// desc for problem definition
constexpr static GemmToConvFwdTransformer dummy_conv_to_gemm_transformer;
constexpr static ConvToGemmFwdTransformer dummy_conv_to_gemm_transformer;
using AGridDesc_M_K =
remove_cvref_t<decltype(MakeAGridDescriptor_M_K<ALayout>(dummy_conv_to_gemm_transformer))>;
using BGridDesc_N_K =
......@@ -617,7 +617,7 @@ struct CodegenDeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
// D batch stride
compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
GemmToConvFwdTransformer conv_to_gemm_transformer_d{a_g_n_c_wis_lengths,
ConvToGemmFwdTransformer conv_to_gemm_transformer_d{a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
......@@ -686,7 +686,7 @@ struct CodegenDeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
// tensor descriptors for problem definiton
index_t num_group_;
GemmToConvFwdTransformer conv_to_gemm_transformer_;
ConvToGemmFwdTransformer conv_to_gemm_transformer_;
AGridDesc_M_K a_grid_desc_m_k_;
BGridDesc_N_K b_grid_desc_n_k_;
......@@ -943,6 +943,77 @@ struct CodegenDeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
b_element_op,
cde_element_op};
}
static __device__ __host__ auto MakeArgument(
APointers p_as,
BPointers p_bs,
const ck::Array<const void*, NumDTensor>& p_ds,
void* p_e,
const ck::Array<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const ck::Array<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const ck::Array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const ck::Array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const ck::Array<ck::Array<long_index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
const ck::Array<ck::Array<long_index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
const ck::Array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const ck::Array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const ck::Array<long_index_t, NDimSpatial>& conv_filter_strides,
const ck::Array<long_index_t, NDimSpatial>& conv_filter_dilations,
const ck::Array<long_index_t, NDimSpatial>& input_left_pads,
const ck::Array<long_index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op)
{
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_i32;
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_dilations_i32;
std::array<index_t, NDimSpatial> input_left_pads_i32;
std::array<index_t, NDimSpatial> input_right_pads_i32;
array_convert(a_g_n_c_wis_lengths_i32, a_g_n_c_wis_lengths);
array_convert(a_g_n_c_wis_strides_i32, a_g_n_c_wis_strides);
array_convert(b_g_k_c_xs_lengths_i32, b_g_k_c_xs_lengths);
array_convert(b_g_k_c_xs_strides_i32, b_g_k_c_xs_strides);
for(index_t d = 0; d < NumDTensor; d++)
{
array_convert(ds_g_n_k_wos_lengths_i32[d], ds_g_n_k_wos_lengths[d]);
array_convert(ds_g_n_k_wos_strides_i32[d], ds_g_n_k_wos_strides[d]);
}
array_convert(e_g_n_k_wos_lengths_i32, e_g_n_k_wos_lengths);
array_convert(e_g_n_k_wos_strides_i32, e_g_n_k_wos_strides);
array_convert(conv_filter_strides_i32, conv_filter_strides);
array_convert(conv_filter_dilations_i32, conv_filter_dilations);
array_convert(input_left_pads_i32, input_left_pads);
array_convert(input_right_pads_i32, input_right_pads);
return Argument{p_as,
p_bs,
p_ds,
p_e,
a_g_n_c_wis_lengths_i32,
a_g_n_c_wis_strides_i32,
b_g_k_c_xs_lengths_i32,
b_g_k_c_xs_strides_i32,
ds_g_n_k_wos_lengths_i32,
ds_g_n_k_wos_strides_i32,
e_g_n_k_wos_lengths_i32,
e_g_n_k_wos_strides_i32,
conv_filter_strides_i32,
conv_filter_dilations_i32,
input_left_pads_i32,
input_right_pads_i32,
a_element_op,
b_element_op,
cde_element_op};
}
};
} // namespace device
......
include/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp
View file @ 0eb75e21
......@@ -64,7 +64,7 @@ struct DeviceColumnToImageImpl
static constexpr auto spatial_offset = Number<3>{};
using GemmToConvFwdTransformer =
using ConvToGemmFwdTransformer =
TransformConvFwdToGemm<NDimSpatial, ConvolutionForwardSpecialization::Default>;
static constexpr auto matrix_padder =
MatrixPadder<GemmSpecialization::MKPadding, index_t, index_t, index_t>{
......@@ -233,7 +233,7 @@ struct DeviceColumnToImageImpl
: independent_filter_stride;
}
GemmToConvFwdTransformer conv_to_gemm_transformer{a_g_n_c_wis_lengths,
ConvToGemmFwdTransformer conv_to_gemm_transformer{a_g_n_c_wis_lengths,
image_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
{}, // not needed for A Descriptor
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp
View file @ 0eb75e21
......@@ -168,15 +168,11 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
// rotating mem
rotating_mem.Next();
// clear c mem
if constexpr(!is_same<remove_cvref_t<CDataType>, bhalf_t>::value)
{
if(arg_.KBatch > 1)
hipGetErrorString(
hipMemsetAsync(arg_.p_c_grid,
hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
0,
arg_.M * arg_.N * sizeof(CDataType),
stream_config.stream_id_));
}
};
ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
......@@ -189,15 +185,12 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
arg_);
}
else
{
if constexpr(!is_same<remove_cvref_t<CDataType>, bhalf_t>::value)
{
if(arg.KBatch > 1)
hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
0,
arg.M * arg.N * sizeof(CDataType),
stream_config.stream_id_));
}
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
......@@ -214,8 +207,6 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
{
if(arg.KBatch > 1)
{
if constexpr(!is_same<remove_cvref_t<CDataType>, bhalf_t>::value)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
......@@ -224,7 +215,6 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
minimum_occupancy>;
Run(kernel);
}
}
else
{
const auto kernel =
......@@ -239,13 +229,11 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
{
if(arg.KBatch > 1)
{
if constexpr(!is_same<remove_cvref_t<CDataType>, bhalf_t>::value)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -255,8 +243,8 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Full)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -266,8 +254,7 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Two)
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
......@@ -326,8 +313,7 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Six)
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
......@@ -354,7 +340,6 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
}
}
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
......@@ -472,8 +457,6 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
{
if(arg.KBatch > 1)
{
if constexpr(!is_same<remove_cvref_t<CDataType>, bhalf_t>::value)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
......@@ -496,7 +479,6 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
Run(kernel);
}
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
......@@ -524,13 +506,11 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
else
{
if(arg.KBatch > 1)
{
if constexpr(!is_same<remove_cvref_t<CDataType>, bhalf_t>::value)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -539,8 +519,8 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
}
else
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -548,7 +528,6 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
Run(kernel);
}
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
......@@ -579,10 +558,7 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
// Tail number always 1
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
{
if(arg.KBatch > 1)
{
if constexpr(!is_same<remove_cvref_t<CDataType>, bhalf_t>::value)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
......@@ -591,7 +567,6 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
minimum_occupancy>;
Run(kernel);
}
}
else
{
const auto kernel =
......@@ -628,6 +603,11 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
return false;
}
if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> && arg.KBatch > 1)
{
return false;
}
if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding ||
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
View file @ 0eb75e21
......@@ -238,14 +238,14 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
using GemmToConvFwdTransformer = TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>;
using ConvToGemmFwdTransformer = TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>;
static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock};
template <typename ALay>
static auto
MakeAGridDescriptor_AK0_M_AK1(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
MakeAGridDescriptor_AK0_M_AK1(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto in_gemmmraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>();
......@@ -266,7 +266,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
template <typename BLay>
static auto
MakeBGridDescriptor_BK0_N_BK1(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
MakeBGridDescriptor_BK0_N_BK1(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto wei_gemmnraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>();
......@@ -287,7 +287,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
}
template <typename ELay>
static auto MakeEGridDescriptor_M_N(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeEGridDescriptor_M_N(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto out_gemmmraw_gemmnraw_desc =
conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>();
......@@ -298,7 +298,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
return out_gemmm_gemmn_desc;
}
static auto MakeDsGridDescriptor_M_N(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeDsGridDescriptor_M_N(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
return generate_tuple(
[&](auto i) {
......@@ -310,7 +310,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
}
// desc for problem definition
constexpr static GemmToConvFwdTransformer dummy_conv_to_gemm_transformer;
constexpr static ConvToGemmFwdTransformer dummy_conv_to_gemm_transformer;
using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(MakeAGridDescriptor_AK0_M_AK1<ALayout>(
dummy_conv_to_gemm_transformer))>;
using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(MakeBGridDescriptor_BK0_N_BK1<BLayout>(
......@@ -447,7 +447,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
GemmToConvFwdTransformer conv_to_gemm_transformer_d{a_g_n_c_wis_lengths,
ConvToGemmFwdTransformer conv_to_gemm_transformer_d{a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
......@@ -511,7 +511,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
// tensor descriptors for problem definiton
index_t num_group_;
GemmToConvFwdTransformer conv_to_gemm_transformer_;
ConvToGemmFwdTransformer conv_to_gemm_transformer_;
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
......@@ -836,6 +836,79 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
cde_element_op};
}
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<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_lengths,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_dilations,
const std::array<long_index_t, NDimSpatial>& input_left_pads,
const std::array<long_index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op)
{
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_i32;
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_dilations_i32;
std::array<index_t, NDimSpatial> input_left_pads_i32;
std::array<index_t, NDimSpatial> input_right_pads_i32;
array_convert(a_g_n_c_wis_lengths_i32, a_g_n_c_wis_lengths);
array_convert(a_g_n_c_wis_strides_i32, a_g_n_c_wis_strides);
array_convert(b_g_k_c_xs_lengths_i32, b_g_k_c_xs_lengths);
array_convert(b_g_k_c_xs_strides_i32, b_g_k_c_xs_strides);
for(index_t d = 0; d < NumDTensor; d++)
{
array_convert(ds_g_n_k_wos_lengths_i32[d], ds_g_n_k_wos_lengths[d]);
array_convert(ds_g_n_k_wos_strides_i32[d], ds_g_n_k_wos_strides[d]);
}
array_convert(e_g_n_k_wos_lengths_i32, e_g_n_k_wos_lengths);
array_convert(e_g_n_k_wos_strides_i32, e_g_n_k_wos_strides);
array_convert(conv_filter_strides_i32, conv_filter_strides);
array_convert(conv_filter_dilations_i32, conv_filter_dilations);
array_convert(input_left_pads_i32, input_left_pads);
array_convert(input_right_pads_i32, input_right_pads);
return Argument{p_a,
p_b,
p_ds,
p_e,
a_g_n_c_wis_lengths_i32,
a_g_n_c_wis_strides_i32,
b_g_k_c_xs_lengths_i32,
b_g_k_c_xs_strides_i32,
ds_g_n_k_wos_lengths_i32,
ds_g_n_k_wos_strides_i32,
e_g_n_k_wos_lengths_i32,
e_g_n_k_wos_strides_i32,
conv_filter_strides_i32,
conv_filter_dilations_i32,
input_left_pads_i32,
input_right_pads_i32,
a_element_op,
b_element_op,
cde_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseArgument> MakeArgumentPointer(
......@@ -880,6 +953,79 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
cde_element_op);
}
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<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_lengths,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_dilations,
const std::array<long_index_t, NDimSpatial>& input_left_pads,
const std::array<long_index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) override
{
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_i32;
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_dilations_i32;
std::array<index_t, NDimSpatial> input_left_pads_i32;
std::array<index_t, NDimSpatial> input_right_pads_i32;
array_convert(a_g_n_c_wis_lengths_i32, a_g_n_c_wis_lengths);
array_convert(a_g_n_c_wis_strides_i32, a_g_n_c_wis_strides);
array_convert(b_g_k_c_xs_lengths_i32, b_g_k_c_xs_lengths);
array_convert(b_g_k_c_xs_strides_i32, b_g_k_c_xs_strides);
for(index_t d = 0; d < NumDTensor; d++)
{
array_convert(ds_g_n_k_wos_lengths_i32[d], ds_g_n_k_wos_lengths[d]);
array_convert(ds_g_n_k_wos_strides_i32[d], ds_g_n_k_wos_strides[d]);
}
array_convert(e_g_n_k_wos_lengths_i32, e_g_n_k_wos_lengths);
array_convert(e_g_n_k_wos_strides_i32, e_g_n_k_wos_strides);
array_convert(conv_filter_strides_i32, conv_filter_strides);
array_convert(conv_filter_dilations_i32, conv_filter_dilations);
array_convert(input_left_pads_i32, input_left_pads);
array_convert(input_right_pads_i32, input_right_pads);
return std::make_unique<Argument>(p_a,
p_b,
p_ds,
p_e,
a_g_n_c_wis_lengths_i32,
a_g_n_c_wis_strides_i32,
b_g_k_c_xs_lengths_i32,
b_g_k_c_xs_strides_i32,
ds_g_n_k_wos_lengths_i32,
ds_g_n_k_wos_strides_i32,
e_g_n_k_wos_lengths_i32,
e_g_n_k_wos_strides_i32,
conv_filter_strides_i32,
conv_filter_dilations_i32,
input_left_pads_i32,
input_right_pads_i32,
a_element_op,
b_element_op,
cde_element_op);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
View file @ 0eb75e21
......@@ -234,14 +234,14 @@ struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimS
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
using GemmToConvFwdTransformer = TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>;
using ConvToGemmFwdTransformer = TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>;
static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock};
template <typename ALay>
static auto
MakeAGridDescriptor_AK0_M_AK1(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
MakeAGridDescriptor_AK0_M_AK1(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto in_gemmmraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>();
......@@ -263,7 +263,7 @@ struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimS
template <typename BLay>
static auto
MakeBGridDescriptor_BK0_N_BK1(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
MakeBGridDescriptor_BK0_N_BK1(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto wei_gemmnraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>();
......@@ -284,7 +284,7 @@ struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimS
}
template <typename CLay>
static auto MakeCGridDescriptor_M_N(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeCGridDescriptor_M_N(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto out_gemmmraw_gemmnraw_desc =
conv_to_gemm_transformer.template MakeCDescriptor_M_N<CLay>();
......@@ -296,7 +296,7 @@ struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimS
}
// desc for problem definition
constexpr static GemmToConvFwdTransformer dummy_conv_to_gemm_transformer;
constexpr static ConvToGemmFwdTransformer dummy_conv_to_gemm_transformer;
using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(MakeAGridDescriptor_AK0_M_AK1<ALayout>(
dummy_conv_to_gemm_transformer))>;
using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(MakeBGridDescriptor_BK0_N_BK1<BLayout>(
......@@ -452,7 +452,7 @@ struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimS
// tensor descriptors for problem definiton
index_t num_group_;
GemmToConvFwdTransformer conv_to_gemm_transformer_;
ConvToGemmFwdTransformer conv_to_gemm_transformer_;
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp
View file @ 0eb75e21
......@@ -293,7 +293,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
using GemmToConvFwdTransformer = TransformConvFwdToGemm<NDimSpatial,
using ConvToGemmFwdTransformer = TransformConvFwdToGemm<NDimSpatial,
ConvForwardSpecialization,
true /*SplitN*/,
ADataType,
......@@ -304,7 +304,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3
template <typename ALay>
static auto
MakeAGridDescriptor_AK0_M_AK1(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
MakeAGridDescriptor_AK0_M_AK1(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto in_gemmmraw_gemmkraw_desc =
......@@ -327,7 +327,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3
template <typename BLay>
static auto
MakeBGridDescriptor_BK0_N_BK1(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
MakeBGridDescriptor_BK0_N_BK1(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto wei_gemmnraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>();
......@@ -348,7 +348,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3
}
template <typename ELay>
static auto MakeEGridDescriptor_M_N(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeEGridDescriptor_M_N(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto out_gemmmraw_gemmnraw_desc =
......@@ -361,7 +361,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3
}
// desc for problem definition
constexpr static GemmToConvFwdTransformer dummy_conv_to_gemm_transformer;
constexpr static ConvToGemmFwdTransformer dummy_conv_to_gemm_transformer;
using EGridDesc_M_N =
remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>(dummy_conv_to_gemm_transformer))>;
......@@ -495,7 +495,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3
// tensor descriptors for problem definiton
index_t num_group_;
GemmToConvFwdTransformer conv_to_gemm_transformer_;
ConvToGemmFwdTransformer conv_to_gemm_transformer_;
index_t conv_N_per_block_;
......@@ -978,6 +978,12 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3
return false;
}
// Gridwise gemm v3 doesn't verify descriptors size
if(!arg.conv_to_gemm_transformer_.AreDescriptorsSmallerThan2GB())
{
return false;
}
// check Gridwise GEMM
const index_t GemmM = arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
const index_t GemmN = arg.b_grid_desc_bk0_n_bk1_.GetLength(I1);
......@@ -1037,6 +1043,79 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3
cde_element_op};
}
static auto
MakeArgument(const void* p_as,
const void* p_bs,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
const std::array<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_lengths,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_dilations,
const std::array<long_index_t, NDimSpatial>& input_left_pads,
const std::array<long_index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op)
{
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_i32;
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_dilations_i32;
std::array<index_t, NDimSpatial> input_left_pads_i32;
std::array<index_t, NDimSpatial> input_right_pads_i32;
array_convert(a_g_n_c_wis_lengths_i32, a_g_n_c_wis_lengths);
array_convert(a_g_n_c_wis_strides_i32, a_g_n_c_wis_strides);
array_convert(b_g_k_c_xs_lengths_i32, b_g_k_c_xs_lengths);
array_convert(b_g_k_c_xs_strides_i32, b_g_k_c_xs_strides);
for(index_t d = 0; d < NumDTensor; d++)
{
array_convert(ds_g_n_k_wos_lengths_i32[d], ds_g_n_k_wos_lengths[d]);
array_convert(ds_g_n_k_wos_strides_i32[d], ds_g_n_k_wos_strides[d]);
}
array_convert(e_g_n_k_wos_lengths_i32, e_g_n_k_wos_lengths);
array_convert(e_g_n_k_wos_strides_i32, e_g_n_k_wos_strides);
array_convert(conv_filter_strides_i32, conv_filter_strides);
array_convert(conv_filter_dilations_i32, conv_filter_dilations);
array_convert(input_left_pads_i32, input_left_pads);
array_convert(input_right_pads_i32, input_right_pads);
return Argument{p_as,
p_bs,
p_ds,
p_e,
a_g_n_c_wis_lengths_i32,
a_g_n_c_wis_strides_i32,
b_g_k_c_xs_lengths_i32,
b_g_k_c_xs_strides_i32,
ds_g_n_k_wos_lengths_i32,
ds_g_n_k_wos_strides_i32,
e_g_n_k_wos_lengths_i32,
e_g_n_k_wos_strides_i32,
conv_filter_strides_i32,
conv_filter_dilations_i32,
input_left_pads_i32,
input_right_pads_i32,
a_element_op,
b_element_op,
cde_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseArgument> MakeArgumentPointer(
......@@ -1081,6 +1160,79 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3
cde_element_op);
}
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<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_lengths,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_dilations,
const std::array<long_index_t, NDimSpatial>& input_left_pads,
const std::array<long_index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) override
{
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_i32;
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_dilations_i32;
std::array<index_t, NDimSpatial> input_left_pads_i32;
std::array<index_t, NDimSpatial> input_right_pads_i32;
array_convert(a_g_n_c_wis_lengths_i32, a_g_n_c_wis_lengths);
array_convert(a_g_n_c_wis_strides_i32, a_g_n_c_wis_strides);
array_convert(b_g_k_c_xs_lengths_i32, b_g_k_c_xs_lengths);
array_convert(b_g_k_c_xs_strides_i32, b_g_k_c_xs_strides);
for(index_t d = 0; d < NumDTensor; d++)
{
array_convert(ds_g_n_k_wos_lengths_i32[d], ds_g_n_k_wos_lengths[d]);
array_convert(ds_g_n_k_wos_strides_i32[d], ds_g_n_k_wos_strides[d]);
}
array_convert(e_g_n_k_wos_lengths_i32, e_g_n_k_wos_lengths);
array_convert(e_g_n_k_wos_strides_i32, e_g_n_k_wos_strides);
array_convert(conv_filter_strides_i32, conv_filter_strides);
array_convert(conv_filter_dilations_i32, conv_filter_dilations);
array_convert(input_left_pads_i32, input_left_pads);
array_convert(input_right_pads_i32, input_right_pads);
return std::make_unique<Argument>(p_a,
p_b,
p_ds,
p_e,
a_g_n_c_wis_lengths_i32,
a_g_n_c_wis_strides_i32,
b_g_k_c_xs_lengths_i32,
b_g_k_c_xs_strides_i32,
ds_g_n_k_wos_lengths_i32,
ds_g_n_k_wos_strides_i32,
e_g_n_k_wos_lengths_i32,
e_g_n_k_wos_strides_i32,
conv_filter_strides_i32,
conv_filter_dilations_i32,
input_left_pads_i32,
input_right_pads_i32,
a_element_op,
b_element_op,
cde_element_op);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
View file @ 0eb75e21
......@@ -309,13 +309,13 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
using GemmToConvFwdTransformer = TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>;
using ConvToGemmFwdTransformer = 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 GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeAGridDescriptor_M_K(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto in_gemmmraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>();
......@@ -327,7 +327,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
}
template <typename BLay>
static auto MakeBGridDescriptor_N_K(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeBGridDescriptor_N_K(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto wei_gemmnraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>();
......@@ -339,7 +339,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
}
template <typename ELay>
static auto MakeEGridDescriptor_M_N(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeEGridDescriptor_M_N(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto out_gemmmraw_gemmnraw_desc =
conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>();
......@@ -420,7 +420,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
return GetPaddedRGridDescriptor(r_grid_desc_mraw, NHoWo);
}
constexpr static GemmToConvFwdTransformer dummy_conv_to_gemm_transformer;
constexpr static ConvToGemmFwdTransformer dummy_conv_to_gemm_transformer;
using AGridDesc_M_K =
remove_cvref_t<decltype(MakeAGridDescriptor_M_K<ALayout>(dummy_conv_to_gemm_transformer))>;
using BGridDesc_N_K =
......@@ -599,7 +599,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
// D batch stride
compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
GemmToConvFwdTransformer conv_to_gemm_transformer_d{a_g_n_c_wis_lengths,
ConvToGemmFwdTransformer conv_to_gemm_transformer_d{a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
......@@ -649,7 +649,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
EDataType* p_e_grid_;
typename GridwiseGemm::RsGridPointer p_rs_grid_;
GemmToConvFwdTransformer conv_to_gemm_transformer_;
ConvToGemmFwdTransformer conv_to_gemm_transformer_;
// tensor descriptors for problem definiton
AGridDesc_M_K a_grid_desc_m_k_;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
View file @ 0eb75e21
......@@ -135,13 +135,13 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
static constexpr auto BEnableLds =
BEnableLds_auto || BEnableLds_manu || (NumGemmKPrefetchStage > 1);
using GemmToConvFwdTransformer = TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>;
using ConvToGemmFwdTransformer = 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(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeAGridDescriptor(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto in_gemmmraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>();
......@@ -185,7 +185,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
}
template <typename BLay>
static auto MakeBGridDescriptor(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeBGridDescriptor(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto wei_gemmnraw_gemmkraw_desc =
conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>();
......@@ -229,7 +229,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
}
template <typename ELay>
static auto MakeEGridDescriptor_M_N(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeEGridDescriptor_M_N(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
const auto out_gemmmraw_gemmnraw_desc =
conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>();
......@@ -240,7 +240,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
return out_gemmm_gemmn_desc;
}
static auto MakeDsGridDescriptor_M_N(const GemmToConvFwdTransformer& conv_to_gemm_transformer)
static auto MakeDsGridDescriptor_M_N(const ConvToGemmFwdTransformer& conv_to_gemm_transformer)
{
return generate_tuple(
[&](auto i) {
......@@ -252,7 +252,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
}
// desc for problem definition
constexpr static GemmToConvFwdTransformer dummy_conv_to_gemm_transformer;
constexpr static ConvToGemmFwdTransformer dummy_conv_to_gemm_transformer;
using AGridDesc =
decltype(DeviceOp::MakeAGridDescriptor<ALayout>(dummy_conv_to_gemm_transformer));
using BGridDesc =
......@@ -406,7 +406,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
[&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
GemmToConvFwdTransformer conv_to_gemm_transformer_d{a_g_n_c_wis_lengths,
ConvToGemmFwdTransformer conv_to_gemm_transformer_d{a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
......@@ -448,7 +448,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
// tensor descriptors for problem definiton
index_t num_group_;
GemmToConvFwdTransformer conv_to_gemm_transformer_;
ConvToGemmFwdTransformer conv_to_gemm_transformer_;
DsGridDesc_M_N ds_grid_desc_m_n_;
EGridDesc_M_N e_grid_desc_m_n_;
......@@ -772,6 +772,81 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
cde_element_op};
}
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<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_lengths,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_dilations,
const std::array<long_index_t, NDimSpatial>& input_left_pads,
const std::array<long_index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op)
{
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_i32;
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_dilations_i32;
std::array<index_t, NDimSpatial> input_left_pads_i32;
std::array<index_t, NDimSpatial> input_right_pads_i32;
array_convert(a_g_n_c_wis_lengths_i32, a_g_n_c_wis_lengths);
array_convert(a_g_n_c_wis_strides_i32, a_g_n_c_wis_strides);
array_convert(b_g_k_c_xs_lengths_i32, b_g_k_c_xs_lengths);
array_convert(b_g_k_c_xs_strides_i32, b_g_k_c_xs_strides);
for(index_t d = 0; d < NumDTensor; d++)
{
array_convert(ds_g_n_k_wos_lengths_i32[d], ds_g_n_k_wos_lengths[d]);
array_convert(ds_g_n_k_wos_strides_i32[d], ds_g_n_k_wos_strides[d]);
}
array_convert(e_g_n_k_wos_lengths_i32, e_g_n_k_wos_lengths);
array_convert(e_g_n_k_wos_strides_i32, e_g_n_k_wos_strides);
array_convert(conv_filter_strides_i32, conv_filter_strides);
array_convert(conv_filter_dilations_i32, conv_filter_dilations);
array_convert(input_left_pads_i32, input_left_pads);
array_convert(input_right_pads_i32, input_right_pads);
return Argument{p_a,
p_b,
p_ds,
p_e,
a_g_n_c_wis_lengths_i32,
a_g_n_c_wis_strides_i32,
b_g_k_c_xs_lengths_i32,
b_g_k_c_xs_strides_i32,
ds_g_n_k_wos_lengths_i32,
ds_g_n_k_wos_strides_i32,
e_g_n_k_wos_lengths_i32,
e_g_n_k_wos_strides_i32,
conv_filter_strides_i32,
conv_filter_dilations_i32,
input_left_pads_i32,
input_right_pads_i32,
1,
1,
a_element_op,
b_element_op,
cde_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseArgument> MakeArgumentPointer(
......@@ -818,6 +893,81 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
cde_element_op);
}
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<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<long_index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<long_index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_lengths,
const std::array<std::array<long_index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<long_index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_strides,
const std::array<long_index_t, NDimSpatial>& conv_filter_dilations,
const std::array<long_index_t, NDimSpatial>& input_left_pads,
const std::array<long_index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) override
{
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_i32;
std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_i32;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_i32;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_i32;
std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_strides_i32;
std::array<index_t, NDimSpatial> conv_filter_dilations_i32;
std::array<index_t, NDimSpatial> input_left_pads_i32;
std::array<index_t, NDimSpatial> input_right_pads_i32;
array_convert(a_g_n_c_wis_lengths_i32, a_g_n_c_wis_lengths);
array_convert(a_g_n_c_wis_strides_i32, a_g_n_c_wis_strides);
array_convert(b_g_k_c_xs_lengths_i32, b_g_k_c_xs_lengths);
array_convert(b_g_k_c_xs_strides_i32, b_g_k_c_xs_strides);
for(index_t d = 0; d < NumDTensor; d++)
{
array_convert(ds_g_n_k_wos_lengths_i32[d], ds_g_n_k_wos_lengths[d]);
array_convert(ds_g_n_k_wos_strides_i32[d], ds_g_n_k_wos_strides[d]);
}
array_convert(e_g_n_k_wos_lengths_i32, e_g_n_k_wos_lengths);
array_convert(e_g_n_k_wos_strides_i32, e_g_n_k_wos_strides);
array_convert(conv_filter_strides_i32, conv_filter_strides);
array_convert(conv_filter_dilations_i32, conv_filter_dilations);
array_convert(input_left_pads_i32, input_left_pads);
array_convert(input_right_pads_i32, input_right_pads);
return std::make_unique<Argument>(p_a,
p_b,
p_ds,
p_e,
a_g_n_c_wis_lengths_i32,
a_g_n_c_wis_strides_i32,
b_g_k_c_xs_lengths_i32,
b_g_k_c_xs_strides_i32,
ds_g_n_k_wos_lengths_i32,
ds_g_n_k_wos_strides_i32,
e_g_n_k_wos_lengths_i32,
e_g_n_k_wos_strides_i32,
conv_filter_strides_i32,
conv_filter_dilations_i32,
input_left_pads_i32,
input_right_pads_i32,
1,
1,
a_element_op,
b_element_op,
cde_element_op);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
......
include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp
View file @ 0eb75e21
......@@ -57,7 +57,7 @@ struct DeviceImageToColumnImpl
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
using GemmToConvFwdTransformer =
using ConvToGemmFwdTransformer =
TransformConvFwdToGemm<NDimSpatial, ConvolutionForwardSpecialization::Default>;
static constexpr auto matrix_padder =
......@@ -97,7 +97,7 @@ struct DeviceImageToColumnImpl
b_g_k_c_xs_lengths[I2] = C;
c_g_n_k_wos_lengths[I1] = N;
GemmToConvFwdTransformer conv_to_gemm_transformer{a_g_n_c_wis_lengths,
ConvToGemmFwdTransformer conv_to_gemm_transformer{a_g_n_c_wis_lengths,
image_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
{}, // not needed for A Descriptor
......
include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -19,7 +19,7 @@ namespace device {
template <index_t Rank, int NumReduceDim>
std::pair<long_index_t, long_index_t> get_2d_lengths(const std::vector<index_t>& inLengths)
{
static_assert(Rank <= 6, "bigger Rank size not supported!");
static_assert(Rank <= 12, "bigger Rank size not supported!");
long_index_t invariant_total_length = 1;
long_index_t reduce_total_length = 1;
......@@ -38,7 +38,7 @@ std::pair<long_index_t, long_index_t> get_2d_lengths(const std::vector<index_t>&
template <index_t Rank, int NumReduceDim>
std::pair<long_index_t, long_index_t> get_2d_lengths(const std::array<index_t, Rank>& inLengths)
{
static_assert(Rank <= 6, "bigger Rank size not supported!");
static_assert(Rank <= 12, "bigger Rank size not supported!");
long_index_t invariant_total_length = 1;
long_index_t reduce_total_length = 1;
......
include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -51,7 +51,7 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InDataType,
PropagateNan,
OutputIndex>
{
static_assert(Rank <= 6, "Bigger Rank size is not supported!");
static_assert(Rank <= 12, "Bigger Rank size is not supported!");
static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
"Invalid thread cluster size assignments!");
......
include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -47,7 +47,7 @@ struct DeviceReduceThreadWise : public DeviceReduce<InDataType,
OutputIndex>
{
static_assert(Rank <= 6, "Bigger Rank size is not supported!");
static_assert(Rank <= 12, "Bigger Rank size is not supported!");
static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
(InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
......
include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise_multi_d.hpp
View file @ 0eb75e21
......@@ -45,7 +45,7 @@ struct DeviceReduceThreadWiseMultiD : public DeviceReduceMultiD<InDataType,
OutElementwiseOperation>
{
static_assert(Rank <= 6, "Bigger Rank size is not supported!");
static_assert(Rank <= 12, "Bigger Rank size is not supported!");
static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
(InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
......
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