Skip to content

GitLab

Commit 9d3cdd21 authored by carlushuang's avatar carlushuang
Browse files

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

parents 06914eed 13332998
Hide whitespace changes
Inline Side-by-side
Showing with 458 additions and 40 deletions
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_default_policy.hpp
View file @ 9d3cdd21
...@@ -47,7 +47,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy ...@@ -47,7 +47,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
{ {
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType, using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType, typename Problem::ComputeDataType,
typename Problem::BlockShape>; typename Problem::BlockShape,
Problem::Traits::kFastFDiv>;
return BlockWelford<P_>{}; return BlockWelford<P_>{};
} }
...@@ -57,7 +58,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy ...@@ -57,7 +58,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
{ {
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType, using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType, typename Problem::ComputeDataType,
typename Problem::BlockShape>; typename Problem::BlockShape,
Problem::Traits::kFastFDiv>;
return BlockWelfordSync<P_>{}; return BlockWelfordSync<P_>{};
} }
...@@ -67,7 +69,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy ...@@ -67,7 +69,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
{ {
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType, using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType, typename Problem::ComputeDataType,
typename Problem::BlockShape>; typename Problem::BlockShape,
Problem::Traits::kFastFDiv>;
return BlockWelfordCrossWarpSync<P_>{}; return BlockWelfordCrossWarpSync<P_>{};
} }
...@@ -79,7 +82,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy ...@@ -79,7 +82,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
{ {
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType, using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType, typename Problem::ComputeDataType,
typename Problem::BlockShape>; typename Problem::BlockShape,
Problem::Traits::kFastFDiv>;
using block_welford = BlockWelford<P_>; using block_welford = BlockWelford<P_>;
using x_block_tile = using x_block_tile =
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_one_pass.hpp
View file @ 9d3cdd21
...@@ -36,6 +36,7 @@ struct Layernorm2dFwdPipelineOnePass ...@@ -36,6 +36,7 @@ struct Layernorm2dFwdPipelineOnePass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync; static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockLayernorm2dFwdProblem::kPadM static constexpr bool kPadM = false; // TODO - BlockLayernorm2dFwdProblem::kPadM
static constexpr bool kPadN = Problem::Traits::kPadN; static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr bool kFastFDiv = Problem::Traits::kFastFDiv;
static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd; static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd;
static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant; static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant;
...@@ -125,7 +126,15 @@ struct Layernorm2dFwdPipelineOnePass ...@@ -125,7 +126,15 @@ struct Layernorm2dFwdPipelineOnePass
// compute inv-std // compute inv-std
auto inv_std = tile_elementwise_in( auto inv_std = tile_elementwise_in(
[&](const auto& v_) { [&](const auto& v_) {
return type_convert<ComputeDataType>(1.0f) / (sqrt(v_ + epsilon)); if(kFastFDiv && std::is_same_v<ComputeDataType, float>)
{
return type_convert<ComputeDataType>(1.0f) *
__builtin_amdgcn_rcpf(sqrt(v_ + epsilon));
}
else
{
return type_convert<ComputeDataType>(1.0f) / sqrt(v_ + epsilon);
}
}, },
var); var);
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_traits.hpp
View file @ 9d3cdd21
...@@ -39,6 +39,7 @@ template<> struct Layernorm2dFusedQuantEnumName<Layernorm2dFusedQuantEnum::SMOOT ...@@ -39,6 +39,7 @@ template<> struct Layernorm2dFusedQuantEnumName<Layernorm2dFusedQuantEnum::SMOOT
template <bool kPadN_, template <bool kPadN_,
bool kSaveMeanInvStd_, bool kSaveMeanInvStd_,
bool kFastFDiv_,
bool kTwoPass_, bool kTwoPass_,
Layernorm2dFusedAddEnum kFusedAdd_, Layernorm2dFusedAddEnum kFusedAdd_,
Layernorm2dFusedQuantEnum kFusedQuant_> Layernorm2dFusedQuantEnum kFusedQuant_>
...@@ -46,6 +47,7 @@ struct Layernorm2dFwdTraits ...@@ -46,6 +47,7 @@ struct Layernorm2dFwdTraits
{ {
static constexpr bool kPadN = kPadN_; static constexpr bool kPadN = kPadN_;
static constexpr bool kSaveMeanInvStd = kSaveMeanInvStd_; static constexpr bool kSaveMeanInvStd = kSaveMeanInvStd_;
static constexpr bool kFastFDiv = kFastFDiv_;
static constexpr bool kTwoPass = kTwoPass_; static constexpr bool kTwoPass = kTwoPass_;
static constexpr Layernorm2dFusedAddEnum kFusedAdd = kFusedAdd_; static constexpr Layernorm2dFusedAddEnum kFusedAdd = kFusedAdd_;
static constexpr Layernorm2dFusedQuantEnum kFusedQuant = kFusedQuant_; static constexpr Layernorm2dFusedQuantEnum kFusedQuant = kFusedQuant_;
......
include/ck_tile/ops/moe_sorting.hpp 0 → 100644
View file @ 9d3cdd21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp"
#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_pipeline.hpp"
#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_policy.hpp"
#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_problem.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
include/ck_tile/ops/welford/block/block_welford.hpp
View file @ 9d3cdd21
...@@ -11,9 +11,10 @@ namespace ck_tile { ...@@ -11,9 +11,10 @@ namespace ck_tile {
template <typename Problem_, typename Policy_ = void> template <typename Problem_, typename Policy_ = void>
struct BlockWelford struct BlockWelford
{ {
using Problem = remove_cvref_t<Problem_>; using Problem = remove_cvref_t<Problem_>;
using XDataType = typename Problem::XDataType; using XDataType = typename Problem::XDataType;
using ComputeDataType = typename Problem::ComputeDataType; using ComputeDataType = typename Problem::ComputeDataType;
static constexpr bool kFastFDiv = Problem::kFastFDiv;
CK_TILE_DEVICE constexpr BlockWelford() {} CK_TILE_DEVICE constexpr BlockWelford() {}
...@@ -89,7 +90,8 @@ struct BlockWelford ...@@ -89,7 +90,8 @@ struct BlockWelford
template <typename Problem_, typename Policy_ = void> template <typename Problem_, typename Policy_ = void>
struct BlockWelfordSync struct BlockWelfordSync
{ {
using Problem = remove_cvref_t<Problem_>; using Problem = remove_cvref_t<Problem_>;
static constexpr bool kFastFDiv = Problem::kFastFDiv;
template <typename MeanDistributedTensor_, typename VarDistributedTensor_> template <typename MeanDistributedTensor_, typename VarDistributedTensor_>
CK_TILE_DEVICE void CK_TILE_DEVICE void
...@@ -173,8 +175,9 @@ struct BlockWelfordSync ...@@ -173,8 +175,9 @@ struct BlockWelfordSync
template <typename Problem_, typename Policy_ = void> template <typename Problem_, typename Policy_ = void>
struct BlockWelfordCrossWarpSync struct BlockWelfordCrossWarpSync
{ {
using Problem = remove_cvref_t<Problem_>; using Problem = remove_cvref_t<Problem_>;
using BlockShape = typename Problem::BlockShape; using BlockShape = typename Problem::BlockShape;
static constexpr bool kFastFDiv = Problem::kFastFDiv;
template <typename MeanDistributedTensor_> template <typename MeanDistributedTensor_>
CK_TILE_DEVICE static constexpr index_t GetReduceWarps() CK_TILE_DEVICE static constexpr index_t GetReduceWarps()
...@@ -351,12 +354,23 @@ CK_TILE_DEVICE constexpr index_t block_tile_welford_calculate_max_count(int row_ ...@@ -351,12 +354,23 @@ CK_TILE_DEVICE constexpr index_t block_tile_welford_calculate_max_count(int row_
} }
// Note: this function must be called after all the computation // Note: this function must be called after all the computation
template <typename VarDistributedTensor_> template <typename VarDistributedTensor_, bool FastFdiv_ = false>
CK_TILE_DEVICE constexpr void block_tile_welford_post_scale_var(VarDistributedTensor_& var_tensor, CK_TILE_DEVICE constexpr void block_tile_welford_post_scale_var(VarDistributedTensor_& var_tensor,
int count) int count,
bool_constant<FastFdiv_> = {})
{ {
using DataType = typename VarDistributedTensor_::DataType; using DataType = typename VarDistributedTensor_::DataType;
tile_elementwise_inout([&count](auto& x) { x = x / type_convert<DataType>(count); }, tile_elementwise_inout(
var_tensor); [&count](auto& x) {
if(FastFdiv_ && std::is_same_v<DataType, float>)
{
x = x * __builtin_amdgcn_rcpf(type_convert<DataType>(count));
}
else
{
x = x / type_convert<DataType>(count);
}
},
var_tensor);
} }
} // namespace ck_tile } // namespace ck_tile
include/ck_tile/ops/welford/block/block_welford_problem.hpp
View file @ 9d3cdd21
...@@ -7,12 +7,13 @@ ...@@ -7,12 +7,13 @@
namespace ck_tile { namespace ck_tile {
template <typename XDataType_, typename ComputeDataType_, typename BlockShape_> template <typename XDataType_, typename ComputeDataType_, typename BlockShape_, bool kFastFDiv_>
struct BlockWelfordProblem struct BlockWelfordProblem
{ {
using XDataType = remove_cvref_t<XDataType_>; using XDataType = remove_cvref_t<XDataType_>;
using ComputeDataType = remove_cvref_t<ComputeDataType_>; using ComputeDataType = remove_cvref_t<ComputeDataType_>;
using BlockShape = remove_cvref_t<BlockShape_>; using BlockShape = remove_cvref_t<BlockShape_>;
static constexpr bool kFastFDiv = kFastFDiv_;
}; };
} // namespace ck_tile } // namespace ck_tile
include/ck_tile/ops/welford/thread/thread_welford.hpp
View file @ 9d3cdd21
...@@ -7,25 +7,46 @@ ...@@ -7,25 +7,46 @@
namespace ck_tile { namespace ck_tile {
template <typename T> template <typename T, bool kFastFDiv = false>
CK_TILE_DEVICE void welford_update(T& mean, T& var, T x, int count) CK_TILE_DEVICE void welford_update(T& mean, T& var, T x, int count, bool_constant<kFastFDiv> = {})
{ {
// TODO: check nan? maybe no // TODO: check nan? maybe no
T delta = x - mean; T delta = x - mean;
mean += delta / count; if(kFastFDiv && std::is_same_v<T, float>)
{
mean += delta * __builtin_amdgcn_rcpf(count);
}
else
{
mean += delta / count;
}
T delta2 = x - mean; T delta2 = x - mean;
var += delta * delta2; var += delta * delta2;
} }
template <typename T> template <typename T, bool kFastFDiv = false>
CK_TILE_DEVICE static void CK_TILE_DEVICE static void welford_merge(T& mean_a,
welford_merge(T& mean_a, T& var_a, int& count_a, T mean_b, T var_b, int count_b) T& var_a,
int& count_a,
T mean_b,
T var_b,
int count_b,
bool_constant<kFastFDiv> = {})
{ {
int count = count_a + count_b; int count = count_a + count_b;
T count_ = type_convert<T>(count); T count_ = type_convert<T>(count);
T count_a_ = type_convert<T>(count_a); T count_a_ = type_convert<T>(count_a);
T count_b_ = type_convert<T>(count_b); T count_b_ = type_convert<T>(count_b);
T count_b_over_count = count == 0 ? type_convert<T>(0) : count_b_ / count_; T count_b_over_count;
if(kFastFDiv && std::is_same_v<T, float>)
{
count_b_over_count =
count == 0 ? type_convert<T>(0) : count_b_ * __builtin_amdgcn_rcpf(count_);
}
else
{
count_b_over_count = count == 0 ? type_convert<T>(0) : count_b_ / count_;
}
T delta = mean_b - mean_a; T delta = mean_b - mean_a;
mean_a += delta * count_b_over_count; mean_a += delta * count_b_over_count;
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp
View file @ 9d3cdd21
...@@ -39,7 +39,25 @@ template <ck::index_t NDimSpatial, ...@@ -39,7 +39,25 @@ template <ck::index_t NDimSpatial,
ConvolutionBackwardWeightSpecialization ConvSpec, ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler, BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion> BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances = std::tuple< using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_generic_instances =
std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances = std::tuple<
// clang-format off // clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups| //#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
...@@ -64,7 +82,25 @@ template <ck::index_t NDimSpatial, ...@@ -64,7 +82,25 @@ template <ck::index_t NDimSpatial,
ConvolutionBackwardWeightSpecialization ConvSpec, ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler, BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion> BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_bf16_instances = std::tuple< using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_generic_instances =
std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instances = std::tuple<
// clang-format off // clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups| //#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
...@@ -82,6 +118,24 @@ using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_bf16_instances = st ...@@ -82,6 +118,24 @@ using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_bf16_instances = st
// clang-format on // clang-format on
>; >;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_generic_instances =
std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>
// clang-format on
>;
// NGCHW requires transpose, we use vector loads and stores params for them // NGCHW requires transpose, we use vector loads and stores params for them
template <ck::index_t NDimSpatial, template <ck::index_t NDimSpatial,
typename ALayout, typename ALayout,
...@@ -122,6 +176,24 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances ...@@ -122,6 +176,24 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances
// clang-format on // clang-format on
>; >;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_generic_instances =
std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>
// clang-format on
>;
template <ck::index_t NDimSpatial, template <ck::index_t NDimSpatial,
typename ALayout, typename ALayout,
typename BLayout, typename BLayout,
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
View file @ 9d3cdd21
...@@ -352,6 +352,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -352,6 +352,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances( add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_instances( add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev5_instances( add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev5_instances(
...@@ -375,6 +377,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -375,6 +377,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instances( add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_instances( add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_instances( add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_instances(
...@@ -390,6 +394,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -390,6 +394,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> && is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> &&
is_same_v<ComputeTypeB, half_t>) is_same_v<ComputeTypeB, half_t>)
{ {
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev2_instances( add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev2_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev5_instances( add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev5_instances(
...@@ -403,6 +409,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -403,6 +409,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<ComputeTypeA, ck::bhalf_t> && is_same_v<ComputeTypeA, ck::bhalf_t> &&
is_same_v<ComputeTypeB, ck::bhalf_t>) is_same_v<ComputeTypeB, ck::bhalf_t>)
{ {
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev2_instances( add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev2_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev5_instances( add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev5_instances(
...@@ -464,6 +472,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -464,6 +472,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances( add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_instances( add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_instances( add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_instances(
...@@ -487,6 +497,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -487,6 +497,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances( add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_instances( add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev5_instances( add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev5_instances(
...@@ -511,6 +523,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -511,6 +523,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> && is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> &&
is_same_v<ComputeTypeB, half_t>) is_same_v<ComputeTypeB, half_t>)
{ {
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev2_instances( add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev2_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev5_instances( add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev5_instances(
...@@ -524,6 +538,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -524,6 +538,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<ComputeTypeA, ck::bhalf_t> && is_same_v<ComputeTypeA, ck::bhalf_t> &&
is_same_v<ComputeTypeB, ck::bhalf_t>) is_same_v<ComputeTypeB, ck::bhalf_t>)
{ {
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev2_instances( add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev2_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev5_instances( add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev5_instances(
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_xdl.inc
View file @ 9d3cdd21
...@@ -113,6 +113,18 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_in ...@@ -113,6 +113,18 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_in
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_instances( void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC, NHWGC,
...@@ -136,6 +148,19 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_p ...@@ -136,6 +148,19 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_p
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NGCHW,
GKYXC,
NGKHW,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev2_instances( void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NGCHW, NGCHW,
...@@ -173,6 +198,18 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances( ...@@ -173,6 +198,18 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_instances( void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC, NHWGC,
...@@ -196,6 +233,19 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pi ...@@ -196,6 +233,19 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pi
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NGCHW,
GKYXC,
NGKHW,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev2_instances( void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NGCHW, NGCHW,
...@@ -298,6 +348,18 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16 ...@@ -298,6 +348,18 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_instances( void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC, NDHWGC,
...@@ -321,6 +383,19 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf1 ...@@ -321,6 +383,19 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf1
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NGCDHW,
GKZYXC,
NGKDHW,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev2_instances( void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NGCDHW, NGCDHW,
...@@ -358,6 +433,18 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances ...@@ -358,6 +433,18 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_instances( void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC, NDHWGC,
...@@ -381,6 +468,19 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16 ...@@ -381,6 +468,19 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NGCDHW,
GKZYXC,
NGKDHW,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev2_instances( void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev2_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NGCDHW, NGCDHW,
......
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ 9d3cdd21
...@@ -88,19 +88,19 @@ function(add_instance_library INSTANCE_NAME) ...@@ -88,19 +88,19 @@ function(add_instance_library INSTANCE_NAME)
foreach(source IN LISTS ARGN) foreach(source IN LISTS ARGN)
set(INST_TARGETS ${SUPPORTED_GPU_TARGETS}) set(INST_TARGETS ${SUPPORTED_GPU_TARGETS})
if(source MATCHES "_xdl") if(source MATCHES "_xdl")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201) list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
elseif(source MATCHES "_wmma") elseif(source MATCHES "_wmma")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030) list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
elseif(source MATCHES "mha") elseif(source MATCHES "mha")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201) list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif() endif()
#only build the fp8 gemm instances for gfx908/90a if the build argument is set #only build the fp8 gemm instances for gfx908/90a if the build argument is set
if(NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH) if(NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH)
if(source MATCHES "gemm_xdl_universal" AND source MATCHES "f8") if(source MATCHES "gemm_xdl_universal" AND source MATCHES "f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201) list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif() endif()
if(source MATCHES "gemm_multiply_multiply_f8") if(source MATCHES "gemm_multiply_multiply_f8")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201) list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif() endif()
endif() endif()
set(offload_targets) set(offload_targets)
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
View file @ 9d3cdd21
...@@ -15,6 +15,10 @@ set(GROUPED_CONV2D_BWD_WEIGHT ...@@ -15,6 +15,10 @@ set(GROUPED_CONV2D_BWD_WEIGHT
xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_instance.cpp xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_instance.cpp
xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev2_instance.cpp xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev2_instance.cpp
xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev5_instance.cpp xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev5_instance.cpp
xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev1_instance.cpp
xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp
xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev1_instance.cpp
xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp
) )
if(DL_KERNELS) if(DL_KERNELS)
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp 0 → 100644
View file @ 9d3cdd21
// 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_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_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_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NGCHW,
GKYXC,
NGKHW,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_generic_instances<
2,
NGCHW,
GKYXC,
NGKHW,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp 0 → 100644
View file @ 9d3cdd21
// 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_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_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_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NGCHW,
GKYXC,
NGKHW,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_generic_instances<
2,
NGCHW,
GKYXC,
NGKHW,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev1_instance.cpp 0 → 100644
View file @ 9d3cdd21
// 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_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_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_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_generic_instances<
2,
NHWGC,
GKYXC,
NHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_instance.cpp
View file @ 9d3cdd21
...@@ -25,7 +25,7 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_p ...@@ -25,7 +25,7 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_p
// 1. Default // 1. Default
add_device_operation_instances( add_device_operation_instances(
instances, instances,
device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_bf16_instances< device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instances<
2, 2,
NHWGC, NHWGC,
GKYXC, GKYXC,
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_instance.cpp
View file @ 9d3cdd21
...@@ -25,7 +25,7 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_p ...@@ -25,7 +25,7 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_p
// 1. Default // 1. Default
add_device_operation_instances( add_device_operation_instances(
instances, instances,
device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_bf16_instances< device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instances<
2, 2,
NHWGC, NHWGC,
GKYXC, GKYXC,
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev1_instance.cpp 0 → 100644
View file @ 9d3cdd21
// 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_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_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_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_generic_instances<
2,
NHWGC,
GKYXC,
NHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_instance.cpp
View file @ 9d3cdd21
...@@ -25,7 +25,7 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pi ...@@ -25,7 +25,7 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pi
// 1. Default // 1. Default
add_device_operation_instances( add_device_operation_instances(
instances, instances,
device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances< device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances<
2, 2,
NHWGC, NHWGC,
GKYXC, GKYXC,
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev5_instance.cpp
View file @ 9d3cdd21
...@@ -25,7 +25,7 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pi ...@@ -25,7 +25,7 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pi
// 1. Default // 1. Default
add_device_operation_instances( add_device_operation_instances(
instances, instances,
device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances< device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances<
2, 2,
NHWGC, NHWGC,
GKYXC, GKYXC,
......
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