Skip to content

GitLab

"test/vscode:/vscode.git/clone" did not exist on "b077e8390e036503b25c76846fb7e9c77323b95e"
Commit 7c81aee8 authored by carlushuang's avatar carlushuang
Browse files

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

parents 49c39b51 0c9012fb
Expand all Show whitespace changes
Inline Side-by-side
Showing with 476 additions and 73 deletions
example/ck_tile/02_layernorm2d/README.md
View file @ 7c81aee8
......@@ -69,7 +69,7 @@ args:
```
## limitations
Note that `fquant=2`, `fadd=2`, `prec_sx/prec_sy` other than `fp32` are not by default generated. though our kernel template suppor this. (TBD: add some flag in generate.py) to generate those instance on demand. Beside, N>8192 case will by default using two-pass pipeline, and `-fquant=1/2` are not supported yet.
Note that `fquant=2`, `fadd=2`, `prec_sx/prec_sy` other than `fp32` are not by default generated. Though our kernel template suppor this. (TBD: add some flag in generate.py) to generate those instance on demand. Beside, `N>8192` case will by default using two-pass pipeline, and `-fquant=1/2` are not supported yet. If need suport `N>8192` and `fused+residual+store`, you can use this example together with `12_smoothquant`, to construct layernorm+residual, and smoothquant, 2 kernels for this purpose.
```
# some case
......
example/ck_tile/02_layernorm2d/generate.py
View file @ 7c81aee8
......@@ -202,8 +202,9 @@ float layernorm2d_fwd_(const S& s, A a)
using Default2DEpilogueProblem = ck_tile::Default2DEpilogueProblem<ComputeDataType, YDataType, false, Traits_::kPadN, false>;
using Default2DEpilogue = ck_tile::Default2DEpilogue<Default2DEpilogueProblem>;
using DynamicQuantEpilogueProblem = ck_tile::DynamicQuantEpilogueProblem<ComputeDataType, YScaleDataType, YDataType, typename Traits_::Shape,
ck_tile::DynamicQuantEpilogueTraits<false, Traits_::kPadN, false, true/*max3*/>>;
static constexpr bool UseSmoothInputScale = Traits_::kFusedQuant == 1;
using DynamicQuantEpilogueProblem = ck_tile::DynamicQuantEpilogueProblem<ComputeDataType, XScaleDataType, YScaleDataType, YDataType, typename Traits_::Shape,
ck_tile::DynamicQuantEpilogueTraits<false, Traits_::kPadN, UseSmoothInputScale, false, true/*max3*/>>;
using DynamicQuantEpilogue = ck_tile::DynamicQuantEpilogue<DynamicQuantEpilogueProblem>;
......
include/ck_tile/ops/epilogue/dynamic_quant_epilogue.hpp
View file @ 7c81aee8
......@@ -8,17 +8,23 @@
namespace ck_tile {
template <bool kPadM_, bool kPadN_, bool UseRawStore_ = true, bool UseMax3_ = false>
template <bool kPadM_,
bool kPadN_,
bool UseSmoothInputScale_,
bool UseRawStore_ = true,
bool UseMax3_ = false>
struct DynamicQuantEpilogueTraits
{
static constexpr bool kPadM = kPadM_;
static constexpr bool kPadN = kPadN_;
static constexpr bool UseSmoothInputScale = UseSmoothInputScale_;
static constexpr bool UseRawStore = UseRawStore_;
static constexpr bool UseMax3 = UseMax3_;
};
// this epilogue just store out a M*N matrix, row major
template <typename AccDataType_,
typename XScaleDataType_,
typename YScaleDataType_,
typename ODataType_,
typename BlockShape_,
......@@ -26,17 +32,20 @@ template <typename AccDataType_,
struct DynamicQuantEpilogueProblem
{
using AccDataType = remove_cvref_t<AccDataType_>;
using XScaleDataType = remove_cvref_t<XScaleDataType_>;
using YScaleDataType = remove_cvref_t<YScaleDataType_>;
using ODataType = remove_cvref_t<ODataType_>;
using BlockShape = remove_cvref_t<BlockShape_>; // can consum generic 2d shape
using Traits = remove_cvref_t<Traits_>;
};
// TODO: we should put descriptor creation function into policy
template <typename Problem_, typename Policy_ = void>
struct DynamicQuantEpilogue
{
using Problem = remove_cvref_t<Problem_>;
using AccDataType = remove_cvref_t<typename Problem::AccDataType>;
using XScaleDataType = remove_cvref_t<typename Problem::XScaleDataType>;
using YScaleDataType = remove_cvref_t<typename Problem::YScaleDataType>;
using ODataType = remove_cvref_t<typename Problem::ODataType>;
using BlockShape = remove_cvref_t<typename Problem::BlockShape>;
......@@ -63,6 +72,33 @@ struct DynamicQuantEpilogue
return BlockReduce2dCrossWarpSync<P_>{};
}
CK_TILE_DEVICE static constexpr auto MakeSmoothInputScaleTileDistribution()
{
using S = BlockShape;
#if 0
// don't remove this
// Note that if we set encoding purposely like this, you will result in compile fail
// TODO: x_scale create local-scratch to accept arbitrary acc input (with same length)
return make_static_tile_distribution(
tile_distribution_encoding<
sequence<S::Repeat_M, S::WarpPerBlock_M, S::ThreadPerWarp_M>,
tuple<sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N, S::Vector_N>>,
tuple<sequence<0, 1>, sequence<0, 1>>,
tuple<sequence<1, 1>, sequence<2, 2>>,
sequence<0, 1, 1>,
sequence<0, 0, 3>>{});
#else
return make_static_tile_distribution(
tile_distribution_encoding<
sequence<S::WarpPerBlock_M, S::ThreadPerWarp_M>,
tuple<sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N, S::Vector_N>>,
tuple<sequence<0, 1>, sequence<0, 1>>,
tuple<sequence<0, 1>, sequence<1, 2>>,
sequence<1, 1>,
sequence<0, 3>>{});
#endif
}
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
auto reduce_crosswarp_sync = GetBlockReduce2dCrossWarpSync();
......@@ -71,8 +107,12 @@ struct DynamicQuantEpilogue
// TODO: this function assume store out vector size is the same as OAccTile last dimension size
// how do we fix this ?
template <typename ODramWindowTmp, typename YScaleWindow, typename OAccTile>
template <typename ODramWindowTmp,
typename XScaleWindow,
typename YScaleWindow,
typename OAccTile>
CK_TILE_DEVICE auto operator()(ODramWindowTmp& o_dram_window_tmp,
const XScaleWindow& x_scale_window_,
YScaleWindow& y_scale_window,
const OAccTile& o_acc_tile,
void* smem)
......@@ -80,6 +120,18 @@ struct DynamicQuantEpilogue
auto reduce = GetBlockReduce2d();
auto reduce_sync = GetBlockReduce2dSync();
auto reduce_crosswarp_sync = GetBlockReduce2dCrossWarpSync();
const auto x_scale_window =
make_tile_window(x_scale_window_, MakeSmoothInputScaleTileDistribution());
auto x_scale = load_tile(x_scale_window);
auto o_acc_tmp = o_acc_tile;
sweep_tile(o_acc_tmp, [&](auto idx) {
constexpr auto j_idx = make_tuple(idx[number<1>{}]);
const auto xs_ = type_convert<AccDataType>(x_scale[j_idx]);
o_acc_tmp(idx) = o_acc_tmp(idx) * xs_;
});
const auto f_absmax = [](auto acc_, auto v_0_) { return max(acc_, abs(v_0_)); };
......@@ -87,10 +139,9 @@ struct DynamicQuantEpilogue
constexpr auto y_size_per_row =
OAccTile{}.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(
number<1>{});
// constexpr auto y_size_per_row = OAccTile::get_lengths()[number<1>{}];
if constexpr(UseMax3 && std::is_same_v<AccDataType, float> && y_size_per_row % 2 == 0)
{
// fast max3 implementation
// fast max3+abs implementation
const auto f_max3 = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
......@@ -98,11 +149,11 @@ struct DynamicQuantEpilogue
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
return reduce(o_acc_tile, type_convert<AccDataType>(0), f_max3, sequence<1, 2>{});
return reduce(o_acc_tmp, type_convert<AccDataType>(0), f_max3, sequence<1, 2>{});
}
else
{
return reduce(o_acc_tile, type_convert<AccDataType>(0), f_absmax);
return reduce(o_acc_tmp, type_convert<AccDataType>(0), f_absmax);
}
}();
reduce_sync(row_absmax, f_absmax);
......@@ -117,23 +168,20 @@ struct DynamicQuantEpilogue
store_tile(y_scale_window, cast_tile<YScaleDataType>(y_scale));
auto o_acc_scaled_tile =
make_static_distributed_tensor<AccDataType>(o_acc_tile.get_tile_distribution());
sweep_tile(o_acc_tile, [&](auto idx) {
sweep_tile(o_acc_tmp, [&](auto idx) {
constexpr auto row_id = make_tuple(idx[number<0>{}]);
o_acc_scaled_tile(idx) = o_acc_tile[idx] / y_scale(row_id);
o_acc_tmp(idx) = o_acc_tmp[idx] / y_scale(row_id);
});
// TODO: this is ugly
if constexpr(UseRawStore && (kPadM || kPadN))
{
store_tile_raw(o_dram_window_tmp, cast_tile<ODataType>(o_acc_scaled_tile));
store_tile_raw(o_dram_window_tmp, cast_tile<ODataType>(o_acc_tmp));
buffer_store_fence();
}
else
{
store_tile(o_dram_window_tmp, cast_tile<ODataType>(o_acc_scaled_tile));
store_tile(o_dram_window_tmp, cast_tile<ODataType>(o_acc_tmp));
}
}
};
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_default_policy.hpp
View file @ 7c81aee8
......@@ -45,7 +45,7 @@ struct Layernorm2dFwdPipelineDefaultPolicy
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockWelford()
{
using P_ = BlockWelfordProblem<typename Problem::XDataType,
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
......@@ -55,7 +55,7 @@ struct Layernorm2dFwdPipelineDefaultPolicy
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockWelfordSync()
{
using P_ = BlockWelfordProblem<typename Problem::XDataType,
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
......@@ -65,7 +65,7 @@ struct Layernorm2dFwdPipelineDefaultPolicy
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockWelfordCrossWarpSync()
{
using P_ = BlockWelfordProblem<typename Problem::XDataType,
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
......@@ -77,13 +77,13 @@ struct Layernorm2dFwdPipelineDefaultPolicy
{
if constexpr(Problem::kNeedCrossWarpSync)
{
using P_ = BlockWelfordProblem<typename Problem::XDataType,
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
using block_welford = BlockWelford<P_>;
using x_block_tile =
decltype(make_static_distributed_tensor<typename Problem::XDataType>(
decltype(make_static_distributed_tensor<typename Problem::ComputeDataType>(
MakeXBlockTileDistribution<Problem>()));
using mean_var_block_tile =
decltype(block_welford::template MakeMeanVarBlockTile<x_block_tile>());
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_one_pass.hpp
View file @ 7c81aee8
......@@ -87,12 +87,9 @@ struct Layernorm2dFwdPipelineOnePass
x_residual_window_, Policy::template MakeXBlockTileDistribution<Problem>());
auto y_residual_window = make_tile_window(
y_residual_window_, Policy::template MakeXBlockTileDistribution<Problem>());
const auto x_scale_window = make_tile_window(
x_scale_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>());
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
auto x_scale = load_tile(x_scale_window);
int cur_count = 0;
int max_count =
......@@ -106,21 +103,21 @@ struct Layernorm2dFwdPipelineOnePass
const auto gamma = load_tile(gamma_window);
const auto beta = load_tile(beta_window);
auto acc = cast_tile<ComputeDataType>(x);
if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE ||
kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD)
{
sweep_tile(x_resi, [&](auto idx) {
// compute x = x_resi + x
auto re_ = type_convert<ComputeDataType>(x_resi(idx)) +
type_convert<ComputeDataType>(x(idx));
x(idx) = type_convert<XDataType>(re_);
acc(idx) = type_convert<ComputeDataType>(x_resi(idx)) + acc(idx);
});
if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE)
store_tile(y_residual_window, x);
store_tile(y_residual_window, cast_tile<YResidualDataType>(acc));
}
// compute welford each-thread->cross-lane->cross-warp
auto [mean, var] = block_welford(x, cur_count, max_count);
auto [mean, var] = block_welford(acc, cur_count, max_count);
block_welford_sync(mean, var, cur_count);
block_welford_cross_warp_sync(mean, var, cur_count, smem);
block_tile_welford_post_scale_var(var, cur_count);
......@@ -138,7 +135,7 @@ struct Layernorm2dFwdPipelineOnePass
store_tile(inv_std_window, cast_tile<InvStdDataType>(inv_std));
// layernorm computation
auto ln = make_static_distributed_tensor<ComputeDataType>(x.get_tile_distribution());
auto ln = make_static_distributed_tensor<ComputeDataType>(acc.get_tile_distribution());
sweep_tile(ln, [&, mean_ = mean](auto idx) {
constexpr auto i_idx = make_tuple(idx[number<0>{}]);
constexpr auto j_idx = make_tuple(idx[number<1>{}]);
......@@ -146,26 +143,15 @@ struct Layernorm2dFwdPipelineOnePass
const auto gamma_ = type_convert<ComputeDataType>(gamma[j_idx]);
const auto beta_ = type_convert<ComputeDataType>(beta[j_idx]);
const auto x_ = type_convert<ComputeDataType>(x[idx]);
auto ln_ = (x_ - mean_[i_idx]) * inv_std[i_idx] * gamma_ + beta_;
auto ln_ = (acc[idx] - mean_[i_idx]) * inv_std[i_idx] * gamma_ + beta_;
ln(idx) = ln_;
});
if constexpr(kFusedQuant == Layernorm2dFusedQuantEnum::SMOOTH_DYNAMIC_QUANT)
{
// smooth-quant pre-scale, then run rowwise-quant
sweep_tile(ln, [&](auto idx) {
constexpr auto j_idx = make_tuple(idx[number<1>{}]);
const auto xs_ = type_convert<ComputeDataType>(x_scale[j_idx]);
ln(idx) = ln(idx) * xs_;
});
}
if constexpr(kFusedQuant == Layernorm2dFusedQuantEnum::DYNAMIC_QUANT ||
kFusedQuant == Layernorm2dFusedQuantEnum::SMOOTH_DYNAMIC_QUANT)
{
Epilogue{}(y_window_, y_scale_window, ln, smem);
Epilogue{}(y_window_, x_scale_window_, y_scale_window, ln, smem);
}
else
Epilogue{}(y_window_, ln);
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_two_pass.hpp
View file @ 7c81aee8
......@@ -106,7 +106,7 @@ struct Layernorm2dFwdPipelineTwoPass
auto block_welford_cross_warp_sync =
Policy::template GetBlockWelfordCrossWarpSync<Problem>();
using XTensorType = decltype(load_tile(x_window));
using XTensorType = decltype(cast_tile<ComputeDataType>(load_tile(x_window)));
auto mean = block_welford.template MakeMeanVarBlockTile<XTensorType>();
auto var = block_welford.template MakeMeanVarBlockTile<XTensorType>();
......@@ -117,22 +117,22 @@ struct Layernorm2dFwdPipelineTwoPass
move_tile_window(x_window, {0, Block_N});
move_tile_window(x_residual_window, {0, Block_N});
auto acc = cast_tile<ComputeDataType>(x);
if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE ||
kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD)
{
sweep_tile(x_resi, [&](auto idx) {
// compute x = x_resi + x
auto re_ = type_convert<ComputeDataType>(x_resi(idx)) +
type_convert<ComputeDataType>(x(idx));
x(idx) = type_convert<XDataType>(re_);
acc(idx) = type_convert<ComputeDataType>(x_resi(idx)) + acc(idx);
});
if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE)
{
store_tile(y_residual_window, x);
store_tile(y_residual_window, cast_tile<YResidualDataType>(acc));
move_tile_window(y_residual_window, {0, Block_N});
}
}
block_welford(x, mean, var, cur_count, max_count);
block_welford(acc, mean, var, cur_count, max_count);
}
block_welford_sync(mean, var, cur_count);
......@@ -166,21 +166,21 @@ struct Layernorm2dFwdPipelineTwoPass
{
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
auto acc = cast_tile<ComputeDataType>(x);
if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE ||
kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD)
{
sweep_tile(x_resi, [&](auto idx) {
// compute x = x_resi + x
auto re_ = type_convert<ComputeDataType>(x_resi(idx)) +
type_convert<ComputeDataType>(x(idx));
x(idx) = type_convert<XDataType>(re_);
acc(idx) = type_convert<ComputeDataType>(x_resi(idx)) + acc(idx);
});
}
// load gamma/beta (TODO: support no gamma/beta?)
const auto gamma = load_tile(gamma_window);
const auto beta = load_tile(beta_window);
auto ln = make_static_distributed_tensor<ComputeDataType>(x.get_tile_distribution());
auto ln = make_static_distributed_tensor<ComputeDataType>(acc.get_tile_distribution());
sweep_tile(ln, [&, mean_ = mean](auto idx) {
constexpr auto i_idx = make_tuple(idx[number<0>{}]);
......@@ -189,8 +189,7 @@ struct Layernorm2dFwdPipelineTwoPass
const auto gamma_ = type_convert<ComputeDataType>(gamma[j_idx]);
const auto beta_ = type_convert<ComputeDataType>(beta[j_idx]);
const auto x_ = type_convert<ComputeDataType>(x[idx]);
auto ln_ = (x_ - mean_[i_idx]) * inv_std[i_idx] * gamma_ + beta_;
auto ln_ = (acc(idx) - mean_[i_idx]) * inv_std[i_idx] * gamma_ + beta_;
ln(idx) = ln_;
});
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp
View file @ 7c81aee8
......@@ -131,6 +131,31 @@ using device_grouped_conv_fwd_xdl_f32_comp_instances = std::tuple<
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_int8_comp_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 256, 32, 8, 8, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 256, 32, 8, 8, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 256, 32, 8, 8, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 256, 32, 8, 8, 16, 16, 8, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
// AGPR Spill when use permuted lds layout. so, use padding for these two.
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp
View file @ 7c81aee8
......@@ -87,6 +87,25 @@ using device_grouped_conv_fwd_xdl_large_tensor_f32_instances = std::tuple<
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_large_tensor_int8_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp
View file @ 7c81aee8
......@@ -154,6 +154,43 @@ using device_grouped_conv_fwd_xdl_f32_mem_instances = std::tuple<
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
BlockGemmPipelineScheduler BlkGemmPipeSched>
using device_grouped_conv_fwd_xdl_int8_mem_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 32, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 64, 16, 16, 128, 8, 8, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 16, 32, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
// Memory friendly
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 32, 64, 8, 8, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 16, 64, 8, 8, 16, 16, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 32, 64, 8, 8, 32, 32, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 16, 64, 8, 8, 16, 16, 4, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 32, 64, 8, 8, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 16, 64, 8, 8, 16, 16, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 32, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 64, 16, 16, 128, 8, 8, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 16, 32, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 32, 64, 64, 8, 8, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 16, 128, 64, 8, 8, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 128, 32, 128, 64, 8, 8, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 16, 256, 64, 8, 8, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 32, 256, 64, 8, 8, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp
View file @ 7c81aee8
......@@ -90,6 +90,25 @@ using device_grouped_conv_fwd_xdl_merged_groups_f32_instances = std::tuple<
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_merged_groups_int8_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Instances with NumGroupsPerBatch > 1
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, int8_t, int8_t, LoopScheduler::Default, 8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, int8_t, int8_t, LoopScheduler::Default, 16>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, int8_t, int8_t, int32_t, int8_t, DsLayout, int8_t, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, int8_t, int8_t, LoopScheduler::Default, 32>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View file @ 7c81aee8
......@@ -122,6 +122,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif // DL_KERNELS
#ifdef CK_USE_XDL
// 1D
// layout GNWC/GKXC/GNWK
if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, GNWC> &&
is_same_v<WeiLayout, GKXC> && is_same_v<OutLayout, GNWK>)
{
......@@ -160,7 +162,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
}
#endif
}
// 2D
// layout GNHWC/GKYXC/GNHWK
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
{
......@@ -191,7 +194,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
}
#endif
}
// layout NHWGC/GKYXC/NHWGK
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
{
......@@ -247,8 +250,27 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_INT8
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t> && is_same_v<AComputeType, int8_t> &&
is_same_v<BComputeType, int8_t>)
{
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_int8_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_int8_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instances(
op_ptrs);
}
#endif
}
// layout NGCHW/GKYXC/NGKHW
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NGCHW> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NGKHW>)
{
......@@ -282,8 +304,26 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
op_ptrs);
}
#endif
#ifdef CK_ENABLE_INT8
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t> && is_same_v<AComputeType, int8_t> &&
is_same_v<BComputeType, int8_t>)
{
add_device_grouped_conv2d_fwd_xdl_merged_groups_ngchw_gkyxc_ngkhw_int8_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_comp_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_mem_intra_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_mem_inter_instances(
op_ptrs);
}
#endif
}
// 3D
// layout GNDHWC/GKZYXC/GNDHWK
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, GNDHWK>)
{
......@@ -323,6 +363,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif
}
// layout NDHWGC/GKZYXC/NDHWGK
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
{
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_comp_xdl.inc
View file @ 7c81aee8
......@@ -57,6 +57,22 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instances(
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
// grouped conv2d forward, NGCHW/GKYXC/NGKHW
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f16_comp_instances(
......@@ -90,6 +106,22 @@ void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f32_comp_instances(
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKYXC,
Empty_Tuple,
NGKHW,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_mem_inter_xdl.inc
View file @ 7c81aee8
......@@ -57,6 +57,22 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instances
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
// grouped conv2d forward, NGCHW/GKYXC/NGKHW
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f16_mem_inter_instances(
......@@ -90,6 +106,22 @@ void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f32_mem_inter_instances
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKYXC,
Empty_Tuple,
NGKHW,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instances(
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_mem_intra_xdl.inc
View file @ 7c81aee8
......@@ -57,6 +57,22 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instances
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
// grouped conv2d forward, NGCHW/GKYXC/NGKHW
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f16_mem_intra_instances(
......@@ -90,6 +106,22 @@ void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f32_mem_intra_instances
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKYXC,
Empty_Tuple,
NGKHW,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances(
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_xdl.inc
View file @ 7c81aee8
......@@ -171,6 +171,22 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
// grouped conv2d forward, NGCHW/GKYXC/NGKHW
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f16_instances(
......@@ -204,6 +220,22 @@ void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f32_instances(
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKYXC,
Empty_Tuple,
NGKHW,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, GNDHWC/GKZYXC/GNDHWK
void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_xdl_large_tensor.inc
View file @ 7c81aee8
......@@ -57,6 +57,22 @@ void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instan
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_xdl_merged_groups.inc
View file @ 7c81aee8
......@@ -85,6 +85,36 @@ void add_device_grouped_conv2d_fwd_xdl_merged_groups_ngchw_gkyxc_ngkhw_f32_insta
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_xdl_merged_groups_ngchw_gkyxc_ngkhw_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKYXC,
Empty_Tuple,
NGKHW,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
View file @ 7c81aee8
......@@ -9,45 +9,56 @@ add_instance_library(device_grouped_conv2d_fwd_instance
xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instance.cpp
# NGCHW, GKYXC, NGKHW
xdl/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f16_instance.cpp
xdl/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f32_instance.cpp
xdl/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_instance.cpp
# large tensor
# NHWGC, GKYXC, NHWGK
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instance.cpp
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_int8_instance.cpp
# merged groups
# NHWGC, GKYXC, NHWGK
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instance.cpp
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_int8_instance.cpp
# NGCHW, GKYXC, NGKHW
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_ngchw_gkyxc_ngkhw_f16_instance.cpp
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_ngchw_gkyxc_ngkhw_f32_instance.cpp
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_ngchw_gkyxc_ngkhw_int8_instance.cpp
#mem
# NHWGC, GKYXC, NHWGK
xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp
xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instance.cpp
xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instance.cpp
xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instance.cpp
# NHWGC, GKYXC, NHWGK
xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp
xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_inter_instance.cpp
xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instance.cpp
xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instance.cpp
# NGCHW, GKYXC, NGKHW
xdl/mem/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f16_mem_intra_instance.cpp
xdl/mem/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f32_mem_intra_instance.cpp
xdl/mem/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_mem_intra_instance.cpp
# NGCHW, GKYXC, NGKHW
xdl/mem/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f16_mem_inter_instance.cpp
xdl/mem/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f32_mem_inter_instance.cpp
xdl/mem/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_mem_inter_instance.cpp
#comp
# NHWGC, GKYXC, NHWGK
xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp
xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instance.cpp
xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instance.cpp
xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instance.cpp
# NGCHW, GKYXC, NGKHW
xdl/comp/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f16_comp_instance.cpp
xdl/comp/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_f32_comp_instance.cpp
xdl/comp/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_comp_instance.cpp
#dl
# GNHWC, GKYXC, GNHWK
dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_comp_instance.cpp 0 → 100644
View file @ 7c81aee8
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_fwd_xdl_ngchw_gkyxc_ngkhw_int8_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKYXC,
Empty_Tuple,
NGKHW,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_int8_comp_instances<2,
NGCHW,
GKYXC,
Empty_Tuple,
NGKHW,
ConvFwdDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment