Skip to content

GitLab

"include/vscode:/vscode.git/clone" did not exist on "54d032b06afe442cc883c67eec9eb31588a2e860"
Commit bbd498f1 authored by Harisankar Sadasivan's avatar Harisankar Sadasivan
Browse files

Merge branch 'develop' of... 

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into simple_gemm_dl
parents 86162966 0abc0f87
Hide whitespace changes
Inline Side-by-side
Showing with 685 additions and 280 deletions
include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp
View file @ bbd498f1
...@@ -12,31 +12,42 @@ template <typename GridwiseReduction, ...@@ -12,31 +12,42 @@ template <typename GridwiseReduction,
typename GammaDataType, typename GammaDataType,
typename BetaDataType, typename BetaDataType,
typename YDataType, typename YDataType,
typename SaveMeanInvStdDataType,
typename ComputeDataType, typename ComputeDataType,
typename YElementwiseOperation, typename YElementwiseOperation,
typename GridDesc_M_K> typename GridDesc_M_K,
__global__ void kernel_normalization(const GridDesc_M_K x_grid_desc_m_k, typename GridDesc_M>
const GridDesc_M_K gamma_grid_desc_m_k, __global__ void
const GridDesc_M_K beta_grid_desc_m_k, kernel_normalization(const GridDesc_M_K x_grid_desc_m_k,
const GridDesc_M_K y_grid_desc_m_k, const GridDesc_M_K gamma_grid_desc_m_k,
index_t num_k_block_tile_iteration, const GridDesc_M_K beta_grid_desc_m_k,
ComputeDataType epsilon, const GridDesc_M_K y_grid_desc_m_k,
const XDataType* const __restrict__ p_x_global, const GridDesc_M save_mean_grid_desc_m,
const GammaDataType* const __restrict__ p_gamma_global, const GridDesc_M save_inv_std_grid_desc_m,
const BetaDataType* const __restrict__ p_beta_global, index_t num_k_block_tile_iteration,
YDataType* const __restrict__ p_y_global, ComputeDataType epsilon,
const YElementwiseOperation y_elementwise_op) const XDataType* const __restrict__ p_x_global,
const GammaDataType* const __restrict__ p_gamma_global,
const BetaDataType* const __restrict__ p_beta_global,
YDataType* const __restrict__ p_y_global,
SaveMeanInvStdDataType* const __restrict__ p_save_mean_global,
SaveMeanInvStdDataType* const __restrict__ p_save_inv_std_global,
const YElementwiseOperation y_elementwise_op)
{ {
GridwiseReduction::Run(x_grid_desc_m_k, GridwiseReduction::Run(x_grid_desc_m_k,
gamma_grid_desc_m_k, gamma_grid_desc_m_k,
beta_grid_desc_m_k, beta_grid_desc_m_k,
y_grid_desc_m_k, y_grid_desc_m_k,
save_mean_grid_desc_m,
save_inv_std_grid_desc_m,
num_k_block_tile_iteration, num_k_block_tile_iteration,
epsilon, epsilon,
p_x_global, p_x_global,
p_gamma_global, p_gamma_global,
p_beta_global, p_beta_global,
p_y_global, p_y_global,
p_save_mean_global,
p_save_inv_std_global,
y_elementwise_op); y_elementwise_op);
}; };
...@@ -44,9 +55,11 @@ template <typename XDataType, ...@@ -44,9 +55,11 @@ template <typename XDataType,
typename GammaDataType, typename GammaDataType,
typename BetaDataType, typename BetaDataType,
typename YDataType, typename YDataType,
typename SaveMeanInvStdDataType,
typename ComputeDataType, typename ComputeDataType,
typename YElementwiseOperation, typename YElementwiseOperation,
typename GridDesc_M_K, typename GridDesc_M_K,
typename GridDesc_M,
index_t BlockSize, index_t BlockSize,
index_t MThreadClusterSize, index_t MThreadClusterSize,
index_t KThreadClusterSize, index_t KThreadClusterSize,
...@@ -60,6 +73,7 @@ template <typename XDataType, ...@@ -60,6 +73,7 @@ template <typename XDataType,
index_t BetaSrcVectorSize, index_t BetaSrcVectorSize,
index_t YDstVectorDim, index_t YDstVectorDim,
index_t YDstVectorSize, index_t YDstVectorSize,
index_t SaveMeanInvStdDstVectorSize,
bool UseWelford> bool UseWelford>
auto NormalizationKernelSelector(bool isSweepOnce) auto NormalizationKernelSelector(bool isSweepOnce)
{ {
...@@ -68,9 +82,11 @@ auto NormalizationKernelSelector(bool isSweepOnce) ...@@ -68,9 +82,11 @@ auto NormalizationKernelSelector(bool isSweepOnce)
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
ComputeDataType, ComputeDataType,
YElementwiseOperation, YElementwiseOperation,
GridDesc_M_K, GridDesc_M_K,
GridDesc_M,
BlockSize, BlockSize,
MThreadClusterSize, MThreadClusterSize,
KThreadClusterSize, KThreadClusterSize,
...@@ -84,15 +100,18 @@ auto NormalizationKernelSelector(bool isSweepOnce) ...@@ -84,15 +100,18 @@ auto NormalizationKernelSelector(bool isSweepOnce)
BetaSrcVectorSize, BetaSrcVectorSize,
YDstVectorDim, YDstVectorDim,
YDstVectorSize, YDstVectorSize,
SaveMeanInvStdDstVectorSize,
false>; false>;
using GridwiseNormalizationSweepOnceNaive = using GridwiseNormalizationSweepOnceNaive =
GridwiseNormalizationNaiveVariance_mk_to_mk<XDataType, GridwiseNormalizationNaiveVariance_mk_to_mk<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
ComputeDataType, ComputeDataType,
YElementwiseOperation, YElementwiseOperation,
GridDesc_M_K, GridDesc_M_K,
GridDesc_M,
BlockSize, BlockSize,
MThreadClusterSize, MThreadClusterSize,
KThreadClusterSize, KThreadClusterSize,
...@@ -106,15 +125,18 @@ auto NormalizationKernelSelector(bool isSweepOnce) ...@@ -106,15 +125,18 @@ auto NormalizationKernelSelector(bool isSweepOnce)
BetaSrcVectorSize, BetaSrcVectorSize,
YDstVectorDim, YDstVectorDim,
YDstVectorSize, YDstVectorSize,
SaveMeanInvStdDstVectorSize,
true>; true>;
using GridwiseNormalizationGenericWelford = using GridwiseNormalizationGenericWelford =
GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType, GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
ComputeDataType, ComputeDataType,
YElementwiseOperation, YElementwiseOperation,
GridDesc_M_K, GridDesc_M_K,
GridDesc_M,
BlockSize, BlockSize,
MThreadClusterSize, MThreadClusterSize,
KThreadClusterSize, KThreadClusterSize,
...@@ -128,15 +150,18 @@ auto NormalizationKernelSelector(bool isSweepOnce) ...@@ -128,15 +150,18 @@ auto NormalizationKernelSelector(bool isSweepOnce)
BetaSrcVectorSize, BetaSrcVectorSize,
YDstVectorDim, YDstVectorDim,
YDstVectorSize, YDstVectorSize,
SaveMeanInvStdDstVectorSize,
false>; false>;
using GridwiseNormalizationSweepOnceWelford = using GridwiseNormalizationSweepOnceWelford =
GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType, GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
ComputeDataType, ComputeDataType,
YElementwiseOperation, YElementwiseOperation,
GridDesc_M_K, GridDesc_M_K,
GridDesc_M,
BlockSize, BlockSize,
MThreadClusterSize, MThreadClusterSize,
KThreadClusterSize, KThreadClusterSize,
...@@ -150,6 +175,7 @@ auto NormalizationKernelSelector(bool isSweepOnce) ...@@ -150,6 +175,7 @@ auto NormalizationKernelSelector(bool isSweepOnce)
BetaSrcVectorSize, BetaSrcVectorSize,
YDstVectorDim, YDstVectorDim,
YDstVectorSize, YDstVectorSize,
SaveMeanInvStdDstVectorSize,
true>; true>;
if constexpr(UseWelford) if constexpr(UseWelford)
...@@ -159,17 +185,21 @@ auto NormalizationKernelSelector(bool isSweepOnce) ...@@ -159,17 +185,21 @@ auto NormalizationKernelSelector(bool isSweepOnce)
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
ComputeDataType, ComputeDataType,
YElementwiseOperation, YElementwiseOperation,
GridDesc_M_K> GridDesc_M_K,
GridDesc_M>
: kernel_normalization<GridwiseNormalizationGenericWelford, : kernel_normalization<GridwiseNormalizationGenericWelford,
XDataType, XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
ComputeDataType, ComputeDataType,
YElementwiseOperation, YElementwiseOperation,
GridDesc_M_K>; GridDesc_M_K,
GridDesc_M>;
} }
else else
{ {
...@@ -178,17 +208,21 @@ auto NormalizationKernelSelector(bool isSweepOnce) ...@@ -178,17 +208,21 @@ auto NormalizationKernelSelector(bool isSweepOnce)
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
ComputeDataType, ComputeDataType,
YElementwiseOperation, YElementwiseOperation,
GridDesc_M_K> GridDesc_M_K,
GridDesc_M>
: kernel_normalization<GridwiseNormalizationGenericNaive, : kernel_normalization<GridwiseNormalizationGenericNaive,
XDataType, XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
ComputeDataType, ComputeDataType,
YElementwiseOperation, YElementwiseOperation,
GridDesc_M_K>; GridDesc_M_K,
GridDesc_M>;
} }
} }
......
include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp
View file @ bbd498f1
...@@ -17,11 +17,13 @@ template <typename MeanVarDataType, ...@@ -17,11 +17,13 @@ template <typename MeanVarDataType,
typename GammaDataType, typename GammaDataType,
typename BetaDataType, typename BetaDataType,
typename YDataType, typename YDataType,
typename SaveMeanInvStdDataType,
typename ComputeDataType, typename ComputeDataType,
typename YElementwiseOperation, typename YElementwiseOperation,
typename MeanVarGridDesc_M_KBlock, typename MeanVarGridDesc_M_KBlock,
typename CountGridDesc_M_KBlock, typename CountGridDesc_M_KBlock,
typename XYGammaBetaGridDesc_M_K, typename XYGammaBetaGridDesc_M_K,
typename SaveMeanInvStdGridDesc_M,
index_t BlockSize, index_t BlockSize,
index_t MThreadClusterSize, index_t MThreadClusterSize,
index_t KThreadClusterSize, index_t KThreadClusterSize,
...@@ -34,7 +36,8 @@ template <typename MeanVarDataType, ...@@ -34,7 +36,8 @@ template <typename MeanVarDataType,
index_t BetaSrcVectorDim, index_t BetaSrcVectorDim,
index_t BetaSrcVectorSize, index_t BetaSrcVectorSize,
index_t YDstVectorDim, index_t YDstVectorDim,
index_t YDstVectorSize> index_t YDstVectorSize,
index_t SaveMeanInvStdDstVectorSize>
struct GridwiseNormalizationSplitK2nd struct GridwiseNormalizationSplitK2nd
{ {
static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) || static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
...@@ -45,6 +48,10 @@ struct GridwiseNormalizationSplitK2nd ...@@ -45,6 +48,10 @@ struct GridwiseNormalizationSplitK2nd
(YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0), (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
"Invalid thread slice sizes and/or vector sizes configuration, please check!"); "Invalid thread slice sizes and/or vector sizes configuration, please check!");
static_assert(MThreadSliceSize % SaveMeanInvStdDstVectorSize == 0,
"Invalid thread slice sizes and/or save mean and inverse std vector sizes "
"configuration, please check!");
static_assert(XSrcVectorSize == YDstVectorSize); static_assert(XSrcVectorSize == YDstVectorSize);
static_assert(XSrcVectorSize == GammaSrcVectorSize); static_assert(XSrcVectorSize == GammaSrcVectorSize);
static_assert(XSrcVectorSize == BetaSrcVectorSize); static_assert(XSrcVectorSize == BetaSrcVectorSize);
...@@ -69,6 +76,10 @@ struct GridwiseNormalizationSplitK2nd ...@@ -69,6 +76,10 @@ struct GridwiseNormalizationSplitK2nd
static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed( static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})); make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
using ThreadBufferLengths_M = Sequence<MThreadSliceSize>;
static constexpr auto thread_buffer_desc_m =
make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
using ThreadBufferLengths_M_1 = Sequence<MThreadSliceSize, 1>; using ThreadBufferLengths_M_1 = Sequence<MThreadSliceSize, 1>;
static constexpr auto thread_buffer_desc_m_1 = static constexpr auto thread_buffer_desc_m_1 =
make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, I1)); make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, I1));
...@@ -99,6 +110,8 @@ struct GridwiseNormalizationSplitK2nd ...@@ -99,6 +110,8 @@ struct GridwiseNormalizationSplitK2nd
const XYGammaBetaGridDesc_M_K& gamma_grid_desc_m_k, const XYGammaBetaGridDesc_M_K& gamma_grid_desc_m_k,
const XYGammaBetaGridDesc_M_K& beta_grid_desc_m_k, const XYGammaBetaGridDesc_M_K& beta_grid_desc_m_k,
const XYGammaBetaGridDesc_M_K& y_grid_desc_m_k, const XYGammaBetaGridDesc_M_K& y_grid_desc_m_k,
const SaveMeanInvStdGridDesc_M& save_mean_grid_desc_m,
const SaveMeanInvStdGridDesc_M& save_inv_std_grid_desc_m,
index_t num_k_mean_var_count_iteration, index_t num_k_mean_var_count_iteration,
index_t num_k_block_tile_iteration, index_t num_k_block_tile_iteration,
index_t k_grid_size, index_t k_grid_size,
...@@ -110,6 +123,8 @@ struct GridwiseNormalizationSplitK2nd ...@@ -110,6 +123,8 @@ struct GridwiseNormalizationSplitK2nd
const GammaDataType* const __restrict__ p_gamma_global, const GammaDataType* const __restrict__ p_gamma_global,
const BetaDataType* const __restrict__ p_beta_global, const BetaDataType* const __restrict__ p_beta_global,
YDataType* const __restrict__ p_y_global, YDataType* const __restrict__ p_y_global,
SaveMeanInvStdDataType* const __restrict__ p_save_mean_global,
SaveMeanInvStdDataType* const __restrict__ p_save_inv_std_global,
const YElementwiseOperation y_elementwise_op) const YElementwiseOperation y_elementwise_op)
{ {
// Thread/Block id // Thread/Block id
...@@ -145,6 +160,12 @@ struct GridwiseNormalizationSplitK2nd ...@@ -145,6 +160,12 @@ struct GridwiseNormalizationSplitK2nd
auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_y_global, y_grid_desc_m_k.GetElementSpaceSize()); p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
auto save_mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_save_mean_global, save_mean_grid_desc_m.GetElementSpaceSize());
auto save_inv_std_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_save_inv_std_global, save_inv_std_grid_desc_m.GetElementSpaceSize());
// VGPR // VGPR
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true> StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
in_mean_thread_buf; in_mean_thread_buf;
...@@ -158,6 +179,7 @@ struct GridwiseNormalizationSplitK2nd ...@@ -158,6 +179,7 @@ struct GridwiseNormalizationSplitK2nd
var_thread_buf; var_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true> StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
welford_count_thread_buf; welford_count_thread_buf;
auto& inv_std_thread_buf = var_thread_buf;
auto x_thread_buf = generate_tuple( auto x_thread_buf = generate_tuple(
[&](auto) { [&](auto) {
...@@ -283,6 +305,42 @@ struct GridwiseNormalizationSplitK2nd ...@@ -283,6 +305,42 @@ struct GridwiseNormalizationSplitK2nd
thread_k_cluster_id * YDstVectorSize), thread_k_cluster_id * YDstVectorSize),
y_elementwise_op); y_elementwise_op);
auto threadwise_mean_store =
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
SaveMeanInvStdDataType,
decltype(thread_buffer_desc_m),
SaveMeanInvStdGridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
SaveMeanInvStdDstVectorSize, // ScalarPerVector
InMemoryDataOperationEnum::Set,
1,
true>(
save_mean_grid_desc_m,
make_multi_index(block_m_cluster_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
auto threadwise_inv_std_store =
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
SaveMeanInvStdDataType,
decltype(thread_buffer_desc_m),
SaveMeanInvStdGridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
SaveMeanInvStdDstVectorSize, // ScalarPerVector
InMemoryDataOperationEnum::Set,
1,
true>(
save_inv_std_grid_desc_m,
make_multi_index(block_m_cluster_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
// step1: Merge mean and variance // step1: Merge mean and variance
constexpr auto mean_var_count_thread_copy_step_I0_k = constexpr auto mean_var_count_thread_copy_step_I0_k =
make_multi_index(I0, KThreadClusterSize); make_multi_index(I0, KThreadClusterSize);
...@@ -332,9 +390,33 @@ struct GridwiseNormalizationSplitK2nd ...@@ -332,9 +390,33 @@ struct GridwiseNormalizationSplitK2nd
BlockwiseWelford::Run( BlockwiseWelford::Run(
mean_thread_buf(I), var_thread_buf(I), welford_count_thread_buf(I)); mean_thread_buf(I), var_thread_buf(I), welford_count_thread_buf(I));
inv_std_thread_buf(I) =
type_convert<ComputeDataType>(1.0f) / ck::math::sqrt(var_thread_buf(I) + epsilon);
}); });
// step2: normalization // step2: save mean and inverse std for backward (optional)
if(block_k_cluster_id == 0 && thread_k_cluster_id == 0)
{
if(p_save_mean_global != nullptr)
{
threadwise_mean_store.Run(thread_buffer_desc_m,
make_tuple(I0),
mean_thread_buf,
save_mean_grid_desc_m,
save_mean_global_val_buf);
}
if(p_save_inv_std_global != nullptr)
{
threadwise_inv_std_store.Run(thread_buffer_desc_m,
make_tuple(I0),
inv_std_thread_buf,
save_inv_std_grid_desc_m,
save_inv_std_global_val_buf);
}
}
// step3: normalization
constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize); constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
for(index_t k = 0; k < num_k_block_tile_iteration; ++k) for(index_t k = 0; k < num_k_block_tile_iteration; ++k)
...@@ -360,7 +442,6 @@ struct GridwiseNormalizationSplitK2nd ...@@ -360,7 +442,6 @@ struct GridwiseNormalizationSplitK2nd
}); });
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) { static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) { static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) { static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k = constexpr auto offset_m_k =
...@@ -369,7 +450,7 @@ struct GridwiseNormalizationSplitK2nd ...@@ -369,7 +450,7 @@ struct GridwiseNormalizationSplitK2nd
// normalize // normalize
y_thread_buf(iK0)(Number<offset_m_k>{}) = y_thread_buf(iK0)(Number<offset_m_k>{}) =
(x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) * (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
divisor; inv_std_thread_buf(iM);
// gamma // gamma
y_thread_buf(iK0)(Number<offset_m_k>{}) = y_thread_buf(iK0)(Number<offset_m_k>{}) =
......
include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp
View file @ bbd498f1
...@@ -16,9 +16,11 @@ template <typename XDataType, ...@@ -16,9 +16,11 @@ template <typename XDataType,
typename GammaDataType, typename GammaDataType,
typename BetaDataType, typename BetaDataType,
typename YDataType, typename YDataType,
typename SaveMeanInvStdDataType,
typename ComputeDataType, typename ComputeDataType,
typename YElementwiseOperation, typename YElementwiseOperation,
typename GridDesc_M_K, typename GridDesc_M_K,
typename GridDesc_M,
index_t BlockSize, index_t BlockSize,
index_t MThreadClusterSize, index_t MThreadClusterSize,
index_t KThreadClusterSize, index_t KThreadClusterSize,
...@@ -32,6 +34,7 @@ template <typename XDataType, ...@@ -32,6 +34,7 @@ template <typename XDataType,
index_t BetaSrcVectorSize, index_t BetaSrcVectorSize,
index_t YDstVectorDim, index_t YDstVectorDim,
index_t YDstVectorSize, index_t YDstVectorSize,
index_t SaveMeanInvStdDstVectorSize,
bool SweepOnce> bool SweepOnce>
struct GridwiseNormalizationWelfordVariance_mk_to_mk struct GridwiseNormalizationWelfordVariance_mk_to_mk
{ {
...@@ -43,6 +46,10 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -43,6 +46,10 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
(YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0), (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
"Invalid thread slice sizes and/or vector sizes configuration, please check!"); "Invalid thread slice sizes and/or vector sizes configuration, please check!");
static_assert(MThreadSliceSize % SaveMeanInvStdDstVectorSize == 0,
"Invalid thread slice sizes and/or save mean and inverse std vector sizes "
"configuration, please check!");
static_assert(XSrcVectorSize == YDstVectorSize); static_assert(XSrcVectorSize == YDstVectorSize);
static_assert(XSrcVectorSize == GammaSrcVectorSize); static_assert(XSrcVectorSize == GammaSrcVectorSize);
static_assert(XSrcVectorSize == BetaSrcVectorSize); static_assert(XSrcVectorSize == BetaSrcVectorSize);
...@@ -64,6 +71,10 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -64,6 +71,10 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed( static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})); make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
using ThreadBufferLengths_M = Sequence<MThreadSliceSize>;
static constexpr auto thread_buffer_desc_m =
make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed( using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}))); make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
using ThreadReduceDstDesc_M = using ThreadReduceDstDesc_M =
...@@ -77,6 +88,8 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -77,6 +88,8 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
ThreadClusterLengths_M_K, ThreadClusterLengths_M_K,
ThreadClusterArrangeOrder>; ThreadClusterArrangeOrder>;
using PassThroughOp = tensor_operation::element_wise::PassThrough;
static constexpr auto I0 = Number<0>{}; static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{}; static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{}; static constexpr auto I2 = Number<2>{};
...@@ -114,17 +127,18 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -114,17 +127,18 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
const GridDesc_M_K& gamma_grid_desc_m_k, const GridDesc_M_K& gamma_grid_desc_m_k,
const GridDesc_M_K& beta_grid_desc_m_k, const GridDesc_M_K& beta_grid_desc_m_k,
const GridDesc_M_K& y_grid_desc_m_k, const GridDesc_M_K& y_grid_desc_m_k,
const GridDesc_M& save_mean_grid_desc_m,
const GridDesc_M& save_inv_std_grid_desc_m,
index_t num_k_block_tile_iteration, index_t num_k_block_tile_iteration,
ComputeDataType epsilon, ComputeDataType epsilon,
const XDataType* const __restrict__ p_x_global, const XDataType* const __restrict__ p_x_global,
const GammaDataType* const __restrict__ p_gamma_global, const GammaDataType* const __restrict__ p_gamma_global,
const BetaDataType* const __restrict__ p_beta_global, const BetaDataType* const __restrict__ p_beta_global,
YDataType* const __restrict__ p_y_global, YDataType* const __restrict__ p_y_global,
SaveMeanInvStdDataType* const __restrict__ p_save_mean_global,
SaveMeanInvStdDataType* const __restrict__ p_save_inv_std_global,
const YElementwiseOperation y_elementwise_op) const YElementwiseOperation y_elementwise_op)
{ {
auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
auto x_thread_buf = generate_tuple( auto x_thread_buf = generate_tuple(
[&](auto) { [&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr, return StaticBuffer<AddressSpaceEnum::Vgpr,
...@@ -150,6 +164,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -150,6 +164,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
mean_thread_buf; mean_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true> StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
var_thread_buf; var_thread_buf;
auto& inv_std_thread_buf = var_thread_buf;
const index_t thread_local_id = get_thread_local_1d_id(); const index_t thread_local_id = get_thread_local_1d_id();
const index_t block_global_id = get_block_1d_id(); const index_t block_global_id = get_block_1d_id();
...@@ -226,6 +241,42 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -226,6 +241,42 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
thread_k_cluster_id * YDstVectorSize), thread_k_cluster_id * YDstVectorSize),
y_elementwise_op); y_elementwise_op);
auto threadwise_mean_store =
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
SaveMeanInvStdDataType,
decltype(thread_buffer_desc_m),
GridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
SaveMeanInvStdDstVectorSize, // ScalarPerVector
InMemoryDataOperationEnum::Set,
1,
true>(
save_mean_grid_desc_m,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
auto threadwise_inv_std_store =
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
SaveMeanInvStdDataType,
decltype(thread_buffer_desc_m),
GridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
SaveMeanInvStdDstVectorSize, // ScalarPerVector
InMemoryDataOperationEnum::Set,
1,
true>(
save_inv_std_grid_desc_m,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize); constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
constexpr auto thread_copy_bwd_step_m_k = constexpr auto thread_copy_bwd_step_m_k =
make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize); make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
...@@ -239,6 +290,15 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -239,6 +290,15 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_beta_global, beta_grid_desc_m_k.GetElementSpaceSize()); p_beta_global, beta_grid_desc_m_k.GetElementSpaceSize());
auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
auto save_mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_save_mean_global, save_mean_grid_desc_m.GetElementSpaceSize());
auto save_inv_std_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_save_inv_std_global, save_inv_std_grid_desc_m.GetElementSpaceSize());
auto threadwise_welford = ThreadwiseWelford(); auto threadwise_welford = ThreadwiseWelford();
threadwise_welford.max_count_ = GetKPerThread(x_grid_desc_m_k, thread_k_cluster_id); threadwise_welford.max_count_ = GetKPerThread(x_grid_desc_m_k, thread_k_cluster_id);
...@@ -279,10 +339,33 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -279,10 +339,33 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
int count = threadwise_welford.cur_count_; int count = threadwise_welford.cur_count_;
BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count); BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
inv_std_thread_buf(I) = type_convert<ComputeDataType>(1.0f) /
ck::math::sqrt(var_thread_buf(I) + epsilon);
}); });
// save mean and inverse std for backward (optional)
if(thread_k_cluster_id == 0)
{
if(p_save_mean_global != nullptr)
{
threadwise_mean_store.Run(thread_buffer_desc_m,
make_tuple(I0),
mean_thread_buf,
save_mean_grid_desc_m,
save_mean_global_val_buf);
}
if(p_save_inv_std_global != nullptr)
{
threadwise_inv_std_store.Run(thread_buffer_desc_m,
make_tuple(I0),
inv_std_thread_buf,
save_inv_std_grid_desc_m,
save_inv_std_global_val_buf);
}
}
// normalization
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) { static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) { static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) { static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k = constexpr auto offset_m_k =
...@@ -291,7 +374,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -291,7 +374,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
// normalize // normalize
y_thread_buf(iK0)(Number<offset_m_k>{}) = y_thread_buf(iK0)(Number<offset_m_k>{}) =
(x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) * (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
divisor; inv_std_thread_buf(iM);
// gamma & beta // gamma & beta
y_thread_buf(iK0)(Number<offset_m_k>{}) = y_thread_buf(iK0)(Number<offset_m_k>{}) =
...@@ -360,8 +443,29 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -360,8 +443,29 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
int count = threadwise_welford.cur_count_; int count = threadwise_welford.cur_count_;
BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count); BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
inv_std_thread_buf(I) = 1 / ck::math::sqrt(var_thread_buf(I) + epsilon);
}); });
if(thread_k_cluster_id == 0)
{
if(p_save_mean_global != nullptr)
{
threadwise_mean_store.Run(thread_buffer_desc_m,
make_tuple(I0),
mean_thread_buf,
save_mean_grid_desc_m,
save_mean_global_val_buf);
}
if(p_save_inv_std_global != nullptr)
{
threadwise_inv_std_store.Run(thread_buffer_desc_m,
make_tuple(I0),
inv_std_thread_buf,
save_inv_std_grid_desc_m,
save_inv_std_global_val_buf);
}
}
auto thread_copy_tail_m_k = auto thread_copy_tail_m_k =
(num_k_block_tile_iteration - 1) * ThreadBufferNumber * thread_copy_fwd_step_m_k; (num_k_block_tile_iteration - 1) * ThreadBufferNumber * thread_copy_fwd_step_m_k;
...@@ -393,7 +497,6 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -393,7 +497,6 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
}); });
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) { static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) { static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) { static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k = constexpr auto offset_m_k =
...@@ -402,7 +505,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk ...@@ -402,7 +505,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
// normalize // normalize
y_thread_buf(iK0)(Number<offset_m_k>{}) = y_thread_buf(iK0)(Number<offset_m_k>{}) =
(x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) * (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
divisor; inv_std_thread_buf(iM);
// gamma // gamma
y_thread_buf(iK0)(Number<offset_m_k>{}) = y_thread_buf(iK0)(Number<offset_m_k>{}) =
......
library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
View file @ bbd498f1
...@@ -20,8 +20,9 @@ template <typename XDataType, ...@@ -20,8 +20,9 @@ template <typename XDataType,
typename GammaDataType, typename GammaDataType,
typename BetaDataType, typename BetaDataType,
typename YDataType, typename YDataType,
typename AccDataType, typename SaveMeanInvStdDataType,
typename AccElementwiseOperation> typename ComputeDataType,
typename YElementwiseOperation>
struct ReferenceGroupnorm : public device::BaseOperator struct ReferenceGroupnorm : public device::BaseOperator
{ {
// x = [N, H, W, G, C] // x = [N, H, W, G, C]
...@@ -35,14 +36,18 @@ struct ReferenceGroupnorm : public device::BaseOperator ...@@ -35,14 +36,18 @@ struct ReferenceGroupnorm : public device::BaseOperator
const Tensor<GammaDataType>& gamma, const Tensor<GammaDataType>& gamma,
const Tensor<BetaDataType>& beta, const Tensor<BetaDataType>& beta,
Tensor<YDataType>& y, Tensor<YDataType>& y,
AccElementwiseOperation acc_elementwise_op, Tensor<SaveMeanInvStdDataType>& save_mean,
Tensor<SaveMeanInvStdDataType>& save_inv_std,
YElementwiseOperation y_elementwise_op,
const std::vector<index_t> lengths, const std::vector<index_t> lengths,
AccDataType epsilon) ComputeDataType epsilon)
: x_(x), : x_(x),
gamma_(gamma), gamma_(gamma),
beta_(beta), beta_(beta),
y_(y), y_(y),
acc_elementwise_op_(acc_elementwise_op), save_mean_(save_mean),
save_inv_std_(save_inv_std),
y_elementwise_op_(y_elementwise_op),
lengths_(lengths), lengths_(lengths),
epsilon_(epsilon) epsilon_(epsilon)
{ {
...@@ -52,9 +57,11 @@ struct ReferenceGroupnorm : public device::BaseOperator ...@@ -52,9 +57,11 @@ struct ReferenceGroupnorm : public device::BaseOperator
const Tensor<XDataType> gamma_; const Tensor<XDataType> gamma_;
const Tensor<XDataType> beta_; const Tensor<XDataType> beta_;
Tensor<YDataType>& y_; Tensor<YDataType>& y_;
AccElementwiseOperation acc_elementwise_op_; Tensor<SaveMeanInvStdDataType>& save_mean_;
Tensor<SaveMeanInvStdDataType>& save_inv_std_;
YElementwiseOperation y_elementwise_op_;
std::vector<index_t> lengths_; std::vector<index_t> lengths_;
AccDataType epsilon_; ComputeDataType epsilon_;
}; };
// Invoker // Invoker
...@@ -68,8 +75,8 @@ struct ReferenceGroupnorm : public device::BaseOperator ...@@ -68,8 +75,8 @@ struct ReferenceGroupnorm : public device::BaseOperator
int G = arg.lengths_[3]; int G = arg.lengths_[3];
int C = arg.lengths_[4]; int C = arg.lengths_[4];
Tensor<AccDataType> mean({N, G}); Tensor<ComputeDataType> mean({N, G});
Tensor<AccDataType> var({N, G}); Tensor<ComputeDataType> var({N, G});
// Compute mean & var in [H, W, C] by Welford Algorithm // Compute mean & var in [H, W, C] by Welford Algorithm
// TODO - parallel for each HWC // TODO - parallel for each HWC
...@@ -78,9 +85,9 @@ struct ReferenceGroupnorm : public device::BaseOperator ...@@ -78,9 +85,9 @@ struct ReferenceGroupnorm : public device::BaseOperator
{ {
for(int g = 0; g < G; ++g) for(int g = 0; g < G; ++g)
{ {
AccDataType mean_val = type_convert<AccDataType>(0.0f); ComputeDataType mean_val = type_convert<ComputeDataType>(0.0f);
AccDataType var_val = type_convert<AccDataType>(0.0f); ComputeDataType var_val = type_convert<ComputeDataType>(0.0f);
int32_t curr_count = 0; int32_t curr_count = 0;
for(int h = 0; h < H; ++h) for(int h = 0; h < H; ++h)
{ {
...@@ -89,10 +96,11 @@ struct ReferenceGroupnorm : public device::BaseOperator ...@@ -89,10 +96,11 @@ struct ReferenceGroupnorm : public device::BaseOperator
for(int c = 0; c < C; ++c) for(int c = 0; c < C; ++c)
{ {
curr_count++; curr_count++;
AccDataType x = type_convert<AccDataType>(arg.x_(n, h, w, g, c)); ComputeDataType x =
AccDataType delta = x - mean_val; type_convert<ComputeDataType>(arg.x_(n, h, w, g, c));
ComputeDataType delta = x - mean_val;
mean_val += delta / curr_count; mean_val += delta / curr_count;
AccDataType delta2 = x - mean_val; ComputeDataType delta2 = x - mean_val;
var_val += delta * delta2; var_val += delta * delta2;
} }
} }
...@@ -100,6 +108,12 @@ struct ReferenceGroupnorm : public device::BaseOperator ...@@ -100,6 +108,12 @@ struct ReferenceGroupnorm : public device::BaseOperator
mean(n, g) = mean_val; mean(n, g) = mean_val;
var(n, g) = var_val / curr_count; var(n, g) = var_val / curr_count;
arg.save_mean_(n, g) = ck::type_convert<SaveMeanInvStdDataType>(mean(n, g));
ComputeDataType divisor =
static_cast<ComputeDataType>(1) / ck::math::sqrt(var(n, g) + arg.epsilon_);
arg.save_inv_std_(n, g) = ck::type_convert<SaveMeanInvStdDataType>(divisor);
} }
} }
...@@ -114,15 +128,19 @@ struct ReferenceGroupnorm : public device::BaseOperator ...@@ -114,15 +128,19 @@ struct ReferenceGroupnorm : public device::BaseOperator
{ {
for(int c = 0; c < C; ++c) for(int c = 0; c < C; ++c)
{ {
AccDataType x = type_convert<AccDataType>(arg.x_(n, h, w, g, c)); ComputeDataType x =
AccDataType gamma = type_convert<AccDataType>(arg.gamma_(g, c)); type_convert<ComputeDataType>(arg.x_(n, h, w, g, c));
AccDataType beta = type_convert<AccDataType>(arg.beta_(g, c)); ComputeDataType gamma =
AccDataType mean_val = type_convert<AccDataType>(mean(n, g)); type_convert<ComputeDataType>(arg.gamma_(g, c));
AccDataType var_val = type_convert<AccDataType>(var(n, g)); ComputeDataType beta =
AccDataType y = gamma * (x - mean_val) / type_convert<ComputeDataType>(arg.beta_(g, c));
ck::math::sqrt(arg.epsilon_ + var_val) + ComputeDataType mean_val =
beta; type_convert<ComputeDataType>(mean(n, g));
arg.acc_elementwise_op_(y, y); ComputeDataType var_val = type_convert<ComputeDataType>(var(n, g));
ComputeDataType y = gamma * (x - mean_val) /
ck::math::sqrt(arg.epsilon_ + var_val) +
beta;
arg.y_elementwise_op_(y, y);
arg.y_(n, h, w, g, c) = type_convert<YDataType>(y); arg.y_(n, h, w, g, c) = type_convert<YDataType>(y);
} }
} }
...@@ -159,11 +177,14 @@ struct ReferenceGroupnorm : public device::BaseOperator ...@@ -159,11 +177,14 @@ struct ReferenceGroupnorm : public device::BaseOperator
const Tensor<GammaDataType>& gamma, const Tensor<GammaDataType>& gamma,
const Tensor<BetaDataType>& beta, const Tensor<BetaDataType>& beta,
Tensor<YDataType>& y, Tensor<YDataType>& y,
AccElementwiseOperation acc_elementwise_op, Tensor<SaveMeanInvStdDataType>& save_mean,
Tensor<SaveMeanInvStdDataType>& save_inv_std,
YElementwiseOperation y_elementwise_op,
const std::vector<index_t> lengths, const std::vector<index_t> lengths,
AccDataType epsilon) ComputeDataType epsilon)
{ {
return Argument{x, gamma, beta, y, acc_elementwise_op, lengths, epsilon}; return Argument{
x, gamma, beta, y, save_mean, save_inv_std, y_elementwise_op, lengths, epsilon};
} }
static auto MakeInvoker() { return Invoker{}; } static auto MakeInvoker() { return Invoker{}; }
......
library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
View file @ bbd498f1
...@@ -20,8 +20,9 @@ template <typename XDataType, ...@@ -20,8 +20,9 @@ template <typename XDataType,
typename GammaDataType, typename GammaDataType,
typename BetaDataType, typename BetaDataType,
typename YDataType, typename YDataType,
typename AccDataType, typename SaveMeanInvStdDataType,
typename AccElementwiseOperation, typename ComputeDataType,
typename YElementwiseOperation,
index_t Rank, index_t Rank,
index_t NumReduceDim> index_t NumReduceDim>
struct ReferenceLayernorm : public device::BaseOperator struct ReferenceLayernorm : public device::BaseOperator
...@@ -36,15 +37,19 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -36,15 +37,19 @@ struct ReferenceLayernorm : public device::BaseOperator
const Tensor<GammaDataType>& gamma_n, const Tensor<GammaDataType>& gamma_n,
const Tensor<BetaDataType>& beta_n, const Tensor<BetaDataType>& beta_n,
Tensor<YDataType>& y_m_n, Tensor<YDataType>& y_m_n,
AccElementwiseOperation acc_elementwise_op, Tensor<SaveMeanInvStdDataType>& save_mean_m,
Tensor<SaveMeanInvStdDataType>& save_inv_std_m,
YElementwiseOperation y_elementwise_op,
const std::vector<index_t> lengths, const std::vector<index_t> lengths,
const std::vector<index_t> reduceDims, const std::vector<index_t> reduceDims,
AccDataType epsilon) ComputeDataType epsilon)
: x_m_n_(x_m_n), : x_m_n_(x_m_n),
gamma_n_(gamma_n), gamma_n_(gamma_n),
beta_n_(beta_n), beta_n_(beta_n),
y_m_n_(y_m_n), y_m_n_(y_m_n),
acc_elementwise_op_(acc_elementwise_op), save_mean_m_(save_mean_m),
save_inv_std_m_(save_inv_std_m),
y_elementwise_op_(y_elementwise_op),
lengths_(lengths), lengths_(lengths),
reduceDims_(reduceDims), reduceDims_(reduceDims),
epsilon_(epsilon) epsilon_(epsilon)
...@@ -55,10 +60,12 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -55,10 +60,12 @@ struct ReferenceLayernorm : public device::BaseOperator
const Tensor<XDataType> gamma_n_; const Tensor<XDataType> gamma_n_;
const Tensor<XDataType> beta_n_; const Tensor<XDataType> beta_n_;
Tensor<YDataType>& y_m_n_; Tensor<YDataType>& y_m_n_;
AccElementwiseOperation acc_elementwise_op_; Tensor<SaveMeanInvStdDataType>& save_mean_m_;
Tensor<SaveMeanInvStdDataType>& save_inv_std_m_;
YElementwiseOperation y_elementwise_op_;
std::vector<index_t> lengths_; std::vector<index_t> lengths_;
std::vector<index_t> reduceDims_; std::vector<index_t> reduceDims_;
AccDataType epsilon_; ComputeDataType epsilon_;
}; };
// Invoker // Invoker
...@@ -69,8 +76,8 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -69,8 +76,8 @@ struct ReferenceLayernorm : public device::BaseOperator
int M = arg.lengths_[0]; int M = arg.lengths_[0];
int N = arg.lengths_[1]; int N = arg.lengths_[1];
Tensor<AccDataType> mean({M}); Tensor<ComputeDataType> mean({M});
Tensor<AccDataType> var({M}); Tensor<ComputeDataType> var({M});
for(int m = 0; m < M; ++m) for(int m = 0; m < M; ++m)
{ {
...@@ -79,7 +86,7 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -79,7 +86,7 @@ struct ReferenceLayernorm : public device::BaseOperator
for(int n = 0; n < N; ++n) for(int n = 0; n < N; ++n)
{ {
auto x_val = ck::type_convert<AccDataType>(arg.x_m_n_(m, n)); auto x_val = ck::type_convert<ComputeDataType>(arg.x_m_n_(m, n));
mean(m) += x_val; mean(m) += x_val;
var(m) += x_val * x_val; var(m) += x_val * x_val;
} }
...@@ -90,17 +97,21 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -90,17 +97,21 @@ struct ReferenceLayernorm : public device::BaseOperator
for(int m = 0; m < M; ++m) for(int m = 0; m < M; ++m)
{ {
AccDataType divisor = ComputeDataType divisor =
static_cast<AccDataType>(1) / ck::math::sqrt(var(m) + arg.epsilon_); static_cast<ComputeDataType>(1) / ck::math::sqrt(var(m) + arg.epsilon_);
for(int n = 0; n < N; ++n) for(int n = 0; n < N; ++n)
{ {
auto x_val = ck::type_convert<AccDataType>(arg.x_m_n_(m, n)); auto x_val = ck::type_convert<ComputeDataType>(arg.x_m_n_(m, n));
auto y_val = (x_val - mean(m)) * divisor; auto gamma_val = ck::type_convert<ComputeDataType>(arg.gamma_n_(n));
y_val = (y_val * arg.gamma_n_(n)) + arg.beta_n_(n); auto beta_val = ck::type_convert<ComputeDataType>(arg.beta_n_(n));
arg.acc_elementwise_op_(y_val, y_val); auto y_val = (x_val - mean(m)) * divisor;
y_val = (y_val * gamma_val) + beta_val;
arg.y_elementwise_op_(y_val, y_val);
arg.y_m_n_(m, n) = ck::type_convert<YDataType>(y_val); arg.y_m_n_(m, n) = ck::type_convert<YDataType>(y_val);
} }
arg.save_mean_m_(m) = ck::type_convert<SaveMeanInvStdDataType>(mean(m));
arg.save_inv_std_m_(m) = ck::type_convert<SaveMeanInvStdDataType>(divisor);
} }
return 0; return 0;
...@@ -140,13 +151,23 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -140,13 +151,23 @@ struct ReferenceLayernorm : public device::BaseOperator
const Tensor<GammaDataType>& gamma_n, const Tensor<GammaDataType>& gamma_n,
const Tensor<BetaDataType>& beta_n, const Tensor<BetaDataType>& beta_n,
Tensor<YDataType>& y_m_n, Tensor<YDataType>& y_m_n,
AccElementwiseOperation acc_elementwise_op, Tensor<SaveMeanInvStdDataType>& save_mean_m,
Tensor<SaveMeanInvStdDataType>& save_inv_std_m,
YElementwiseOperation y_elementwise_op,
const std::vector<index_t> lengths, const std::vector<index_t> lengths,
const std::vector<index_t> reduceDims, const std::vector<index_t> reduceDims,
AccDataType epsilon) ComputeDataType epsilon)
{ {
return Argument{ return Argument{x_m_n,
x_m_n, gamma_n, beta_n, y_m_n, acc_elementwise_op, lengths, reduceDims, epsilon}; gamma_n,
beta_n,
y_m_n,
save_mean_m,
save_inv_std_m,
y_elementwise_op,
lengths,
reduceDims,
epsilon};
} }
static auto MakeInvoker() { return Invoker{}; } static auto MakeInvoker() { return Invoker{}; }
......
library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp
View file @ bbd498f1
...@@ -19,13 +19,13 @@ namespace instance { ...@@ -19,13 +19,13 @@ namespace instance {
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
// FP16 // FP16
void add_device_normalization_rank_2_1_f16_instances( void add_device_normalization_rank_2_1_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, PassThrough, 2, 1>>>&); std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, PassThrough, 2, 1>>>&);
void add_device_normalization_rank_4_3_f16_instances( void add_device_normalization_rank_4_3_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, PassThrough, 4, 3>>>&); std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, PassThrough, 4, 3>>>&);
void add_device_normalization_rank_5_3_f16_instances( void add_device_normalization_rank_5_3_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, PassThrough, 5, 3>>>&); std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, PassThrough, 5, 3>>>&);
#endif #endif
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
// FP32 // FP32
...@@ -42,14 +42,15 @@ template <typename XDataType, ...@@ -42,14 +42,15 @@ template <typename XDataType,
typename GammaDataType, typename GammaDataType,
typename BetaDataType, typename BetaDataType,
typename YDataType, typename YDataType,
typename SaveMeanInvStdDataType,
index_t Rank, index_t Rank,
index_t NumReduceDim> index_t NumReduceDim>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormalization< struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormalization<
XDataType, XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
F32,
YDataType, YDataType,
SaveMeanInvStdDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
Rank, Rank,
NumReduceDim>> NumReduceDim>>
...@@ -57,8 +58,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormal ...@@ -57,8 +58,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormal
using DeviceOp = DeviceNormalization<XDataType, using DeviceOp = DeviceNormalization<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
F32,
YDataType, YDataType,
SaveMeanInvStdDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
Rank, Rank,
NumReduceDim>; NumReduceDim>;
...@@ -68,7 +69,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormal ...@@ -68,7 +69,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormal
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F16> && if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F16> &&
is_same_v<BetaDataType, F16> && is_same_v<YDataType, F16>) is_same_v<BetaDataType, F16> && is_same_v<YDataType, F16> &&
is_same_v<SaveMeanInvStdDataType, F32>)
{ {
if constexpr(Rank == 2 && NumReduceDim == 1) if constexpr(Rank == 2 && NumReduceDim == 1)
{ {
...@@ -86,7 +88,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormal ...@@ -86,7 +88,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormal
#endif #endif
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<XDataType, F32> && is_same_v<GammaDataType, F32> && if constexpr(is_same_v<XDataType, F32> && is_same_v<GammaDataType, F32> &&
is_same_v<BetaDataType, F32> && is_same_v<YDataType, F32>) is_same_v<BetaDataType, F32> && is_same_v<YDataType, F32> &&
is_same_v<SaveMeanInvStdDataType, F32>)
{ {
if constexpr(Rank == 2 && NumReduceDim == 1) if constexpr(Rank == 2 && NumReduceDim == 1)
{ {
......
library/include/ck/library/tensor_operation_instance/gpu/normalization_swish.hpp
View file @ bbd498f1
...@@ -19,7 +19,7 @@ namespace instance { ...@@ -19,7 +19,7 @@ namespace instance {
// FP16 // FP16
void add_device_normalization_rank_5_3_swish_f16_instances( void add_device_normalization_rank_5_3_swish_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Swish, 5, 3>>>&); std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Swish, 5, 3>>>&);
// FP32 // FP32
void add_device_normalization_rank_5_3_swish_f32_instances( void add_device_normalization_rank_5_3_swish_f32_instances(
...@@ -27,20 +27,21 @@ void add_device_normalization_rank_5_3_swish_f32_instances( ...@@ -27,20 +27,21 @@ void add_device_normalization_rank_5_3_swish_f32_instances(
// [x, gamma, beta, y] = [f16, f32, f32, f16] // [x, gamma, beta, y] = [f16, f32, f32, f16]
void add_device_normalization_rank_5_3_swish_f16_f32_f32_f16_instances( void add_device_normalization_rank_5_3_swish_f16_f32_f32_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F32, F32, F32, F16, Swish, 5, 3>>>&); std::vector<std::unique_ptr<DeviceNormalization<F16, F32, F32, F16, F32, Swish, 5, 3>>>&);
template <typename XDataType, template <typename XDataType,
typename GammaDataType, typename GammaDataType,
typename BetaDataType, typename BetaDataType,
typename YDataType, typename YDataType,
typename SaveMeanInvStdDataType,
index_t Rank, index_t Rank,
index_t NumReduceDim> index_t NumReduceDim>
struct DeviceOperationInstanceFactory< struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceNormalization<XDataType, ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
F32,
YDataType, YDataType,
SaveMeanInvStdDataType,
ck::tensor_operation::element_wise::Swish, ck::tensor_operation::element_wise::Swish,
Rank, Rank,
NumReduceDim>> NumReduceDim>>
...@@ -48,8 +49,8 @@ struct DeviceOperationInstanceFactory< ...@@ -48,8 +49,8 @@ struct DeviceOperationInstanceFactory<
using DeviceOp = DeviceNormalization<XDataType, using DeviceOp = DeviceNormalization<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
F32,
YDataType, YDataType,
SaveMeanInvStdDataType,
ck::tensor_operation::element_wise::Swish, ck::tensor_operation::element_wise::Swish,
Rank, Rank,
NumReduceDim>; NumReduceDim>;
...@@ -59,7 +60,8 @@ struct DeviceOperationInstanceFactory< ...@@ -59,7 +60,8 @@ struct DeviceOperationInstanceFactory<
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F16> && if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F16> &&
is_same_v<BetaDataType, F16> && is_same_v<YDataType, F16>) is_same_v<BetaDataType, F16> && is_same_v<YDataType, F16> &&
is_same_v<SaveMeanInvStdDataType, F32>)
{ {
if constexpr(Rank == 5 && NumReduceDim == 3) if constexpr(Rank == 5 && NumReduceDim == 3)
{ {
...@@ -67,7 +69,8 @@ struct DeviceOperationInstanceFactory< ...@@ -67,7 +69,8 @@ struct DeviceOperationInstanceFactory<
} }
} }
else if constexpr(is_same_v<XDataType, F32> && is_same_v<GammaDataType, F32> && else if constexpr(is_same_v<XDataType, F32> && is_same_v<GammaDataType, F32> &&
is_same_v<BetaDataType, F32> && is_same_v<YDataType, F32>) is_same_v<BetaDataType, F32> && is_same_v<YDataType, F32> &&
is_same_v<SaveMeanInvStdDataType, F32>)
{ {
if constexpr(Rank == 5 && NumReduceDim == 3) if constexpr(Rank == 5 && NumReduceDim == 3)
{ {
...@@ -75,7 +78,8 @@ struct DeviceOperationInstanceFactory< ...@@ -75,7 +78,8 @@ struct DeviceOperationInstanceFactory<
} }
} }
else if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F32> && else if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F32> &&
is_same_v<BetaDataType, F32> && is_same_v<YDataType, F16>) is_same_v<BetaDataType, F32> && is_same_v<YDataType, F16> &&
is_same_v<SaveMeanInvStdDataType, F32>)
{ {
if constexpr(Rank == 5 && NumReduceDim == 3) if constexpr(Rank == 5 && NumReduceDim == 3)
{ {
......
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp
View file @ bbd498f1
...@@ -11,7 +11,7 @@ namespace instance { ...@@ -11,7 +11,7 @@ namespace instance {
using Pass = ck::tensor_operation::element_wise::PassThrough; using Pass = ck::tensor_operation::element_wise::PassThrough;
void add_device_normalization_rank_5_3_f16_instances( void add_device_normalization_rank_5_3_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 5, 3>>>& std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Pass, 5, 3>>>&
instances) instances)
{ {
add_device_operation_instances(instances, add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp
View file @ bbd498f1
...@@ -11,7 +11,7 @@ namespace instance { ...@@ -11,7 +11,7 @@ namespace instance {
using Swish = ck::tensor_operation::element_wise::Swish; using Swish = ck::tensor_operation::element_wise::Swish;
void add_device_normalization_rank_5_3_swish_f16_f32_f32_f16_instances( void add_device_normalization_rank_5_3_swish_f16_f32_f32_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F32, F32, F32, F16, Swish, 5, 3>>>& std::vector<std::unique_ptr<DeviceNormalization<F16, F32, F32, F16, F32, Swish, 5, 3>>>&
instances) instances)
{ {
add_device_operation_instances( add_device_operation_instances(
......
library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp
View file @ bbd498f1
...@@ -11,7 +11,7 @@ namespace instance { ...@@ -11,7 +11,7 @@ namespace instance {
using Swish = ck::tensor_operation::element_wise::Swish; using Swish = ck::tensor_operation::element_wise::Swish;
void add_device_normalization_rank_5_3_swish_f16_instances( void add_device_normalization_rank_5_3_swish_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Swish, 5, 3>>>& std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Swish, 5, 3>>>&
instances) instances)
{ {
add_device_operation_instances(instances, add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp
View file @ bbd498f1
...@@ -11,7 +11,7 @@ namespace instance { ...@@ -11,7 +11,7 @@ namespace instance {
using Pass = ck::tensor_operation::element_wise::PassThrough; using Pass = ck::tensor_operation::element_wise::PassThrough;
void add_device_normalization_rank_2_1_f16_instances( void add_device_normalization_rank_2_1_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 2, 1>>>& std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Pass, 2, 1>>>&
instances) instances)
{ {
add_device_operation_instances(instances, add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp
View file @ bbd498f1
...@@ -11,7 +11,7 @@ namespace instance { ...@@ -11,7 +11,7 @@ namespace instance {
using Pass = ck::tensor_operation::element_wise::PassThrough; using Pass = ck::tensor_operation::element_wise::PassThrough;
void add_device_normalization_rank_4_3_f16_instances( void add_device_normalization_rank_4_3_f16_instances(
std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 4, 3>>>& std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Pass, 4, 3>>>&
instances) instances)
{ {
add_device_operation_instances(instances, add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp
View file @ bbd498f1
...@@ -22,25 +22,25 @@ template <typename OutElementwise, index_t Rank, index_t Reduce> ...@@ -22,25 +22,25 @@ template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_instances = using device_normalization_f16_instances =
// clang-format off // clang-format off
std::tuple < std::tuple <
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize> // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>, // irregular size DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>, // irregular size
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8> DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
// clang-format on // clang-format on
>; >;
...@@ -48,150 +48,150 @@ template <typename OutElementwise, index_t Rank, index_t Reduce> ...@@ -48,150 +48,150 @@ template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_splitk_f16_instances = using device_normalization_splitk_f16_instances =
// clang-format off // clang-format off
std::tuple < std::tuple <
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize> // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>, // irregular size DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>, DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8> DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_generic_instance = std::tuple< using device_normalization_f16_generic_instance = std::tuple<
// clang-format off // clang-format off
DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1> DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f32_instances = std::tuple< using device_normalization_f32_instances = std::tuple<
// clang-format off // clang-format off
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize> // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4> DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_splitk_f32_instances = std::tuple< using device_normalization_splitk_f32_instances = std::tuple<
// clang-format off // clang-format off
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize> // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4> DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f32_generic_instance = std::tuple< using device_normalization_f32_generic_instance = std::tuple<
// clang-format off // clang-format off
DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1> DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_f32_f32_f16_instances = std::tuple< using device_normalization_f16_f32_f32_f16_instances = std::tuple<
// clang-format off // clang-format off
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize> // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4> DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_splitk_f16_f32_f32_f16_instances = std::tuple< using device_normalization_splitk_f16_f32_f32_f16_instances = std::tuple<
// clang-format off // clang-format off
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize> // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>, DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4> DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_f32_f32_f16_generic_instance = std::tuple< using device_normalization_f16_f32_f32_f16_generic_instance = std::tuple<
// clang-format off // clang-format off
DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1> DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
// clang-format on // clang-format on
>; >;
......
profiler/README.md
View file @ bbd498f1
...@@ -22,7 +22,7 @@ c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1} ...@@ -22,7 +22,7 @@ c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
Best Perf: 1.1933 ms, 107.977 TFlops, 79.0848 GB/s Best Perf: 1.1933 ms, 107.977 TFlops, 79.0848 GB/s
``` ```
## Profile 2d forward convolution kernels ## Profile 2D forward convolution kernels
```bash ```bash
#arg1: tensor operation (conv=Convolution) #arg1: tensor operation (conv=Convolution)
#arg2: data type (0=fp32, 1=fp16) #arg2: data type (0=fp32, 1=fp16)
...@@ -115,7 +115,7 @@ Best Perf: 58.0306 ms, 37.8942 TFlops, 27.7545 GB/s ...@@ -115,7 +115,7 @@ Best Perf: 58.0306 ms, 37.8942 TFlops, 27.7545 GB/s
# arg6: print tensor value (0: no; 1: yes) # arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0: no, 1: yes) # arg7: time kernel (0: no, 1: yes)
# Following arguments (depending on number of spatial dims): # Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d) # Number of spatial dimensions (1=Conv1D, 2=Conv2D, 3=Conv3D)
# G, N, K, C, # G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D) # <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D) # <input image spatial dimensions>, (ie Hi, Wi for 2D)
...@@ -158,7 +158,7 @@ GB/s: 127.947 ...@@ -158,7 +158,7 @@ GB/s: 127.947
# arg6: print tensor value (0: no; 1: yes) # arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0: no, 1: yes) # arg7: time kernel (0: no, 1: yes)
# Following arguments (depending on number of spatial dims): # Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d) # Number of spatial dimensions (1=Conv1D, 2=Conv2D, 3=Conv3D)
# G, N, K, C, # G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D) # <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D) # <input image spatial dimensions>, (ie Hi, Wi for 2D)
...@@ -201,7 +201,7 @@ Note: This kernel use atomic add, this will cause output buffer to be accumulate ...@@ -201,7 +201,7 @@ Note: This kernel use atomic add, this will cause output buffer to be accumulate
# arg7: time kernel (0: no, 1: yes) # arg7: time kernel (0: no, 1: yes)
# arg8: operation type (0: ImageToColumn, 1: ColumnToImage) # arg8: operation type (0: ImageToColumn, 1: ColumnToImage)
# Following arguments (depending on number of spatial dims): # Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d) # Number of spatial dimensions (1=Conv1D, 2=Conv2D, 3=Conv3D)
# G, N, K, C, # G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D) # <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D) # <input image spatial dimensions>, (ie Hi, Wi for 2D)
......
profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
View file @ bbd498f1
...@@ -80,6 +80,8 @@ bool profile_elementwise_layernorm_impl(int do_verification, ...@@ -80,6 +80,8 @@ bool profile_elementwise_layernorm_impl(int do_verification,
Tensor<BetaDataType> beta(gammaBetaLength); Tensor<BetaDataType> beta(gammaBetaLength);
Tensor<YDataType> y(length); Tensor<YDataType> y(length);
Tensor<YDataType> host_y(length); Tensor<YDataType> host_y(length);
Tensor<AccDataType> host_save_mean({M});
Tensor<AccDataType> host_save_inv_std({M});
switch(init_method) switch(init_method)
{ {
...@@ -152,14 +154,23 @@ bool profile_elementwise_layernorm_impl(int do_verification, ...@@ -152,14 +154,23 @@ bool profile_elementwise_layernorm_impl(int do_verification,
BetaDataType, BetaDataType,
YDataType, YDataType,
AccDataType, AccDataType,
AccDataType,
PassThrough, PassThrough,
Rank, Rank,
NumReduceDim>; NumReduceDim>;
ReferenceInstance ref; ReferenceInstance ref;
auto ref_argument = auto ref_argument = ref.MakeArgument(x,
ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, {M, N}, {1}, 1e-4); gamma,
auto ref_invoker = ref.MakeInvoker(); beta,
host_y,
host_save_mean,
host_save_inv_std,
PassThrough{},
{M, N},
{1},
1e-4);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
} }
......
profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp
View file @ bbd498f1
...@@ -66,12 +66,15 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n, ...@@ -66,12 +66,15 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
BetaDataType, BetaDataType,
HDataType, HDataType,
AccDataType, AccDataType,
AccDataType,
HElementOp, HElementOp,
2, 2,
1>; 1>;
Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N}); Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N});
Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N}); Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
Tensor<AccDataType> save_mean({M});
Tensor<AccDataType> save_inv_std({M});
auto ref_gemm = ReferenceGemm{}; auto ref_gemm = ReferenceGemm{};
auto ref_gemm_invoker = ref_gemm.MakeInvoker(); auto ref_gemm_invoker = ref_gemm.MakeInvoker();
...@@ -97,7 +100,7 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n, ...@@ -97,7 +100,7 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
auto ref_layernorm_invoker = ref_layernorm.MakeInvoker(); auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
auto ref_layernorm_argument = ref_layernorm.MakeArgument( auto ref_layernorm_argument = ref_layernorm.MakeArgument(
e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon); e_m_n, gamma_n, beta_n, h_m_n, save_mean, save_inv_std, h_element_op, {M, N}, {1}, epsilon);
ref_layernorm_invoker.Run(ref_layernorm_argument); ref_layernorm_invoker.Run(ref_layernorm_argument);
} }
......
profiler/include/profiler/profile_groupnorm_impl.hpp
View file @ bbd498f1
...@@ -21,8 +21,10 @@ namespace profiler { ...@@ -21,8 +21,10 @@ namespace profiler {
template <typename XDataType, template <typename XDataType,
typename GammaDataType, typename GammaDataType,
typename BetaDataType, typename BetaDataType,
typename AccDataType, typename ComputeDataType,
typename YDataType> typename YDataType,
typename SaveMeanInvStdDataType,
bool SaveMeanInvStd>
bool profile_groupnorm_impl(int do_verification, bool profile_groupnorm_impl(int do_verification,
int init_method, int init_method,
bool do_log, bool do_log,
...@@ -34,6 +36,7 @@ bool profile_groupnorm_impl(int do_verification, ...@@ -34,6 +36,7 @@ bool profile_groupnorm_impl(int do_verification,
if(length.size() != 5) if(length.size() != 5)
return false; return false;
index_t N = length[0];
index_t G = length[3]; index_t G = length[3];
index_t C = length[4]; index_t C = length[4];
...@@ -45,7 +48,14 @@ bool profile_groupnorm_impl(int do_verification, ...@@ -45,7 +48,14 @@ bool profile_groupnorm_impl(int do_verification,
Tensor<GammaDataType> gamma(gammaBetaLength); Tensor<GammaDataType> gamma(gammaBetaLength);
Tensor<BetaDataType> beta(gammaBetaLength); Tensor<BetaDataType> beta(gammaBetaLength);
Tensor<YDataType> y(length); Tensor<YDataType> y(length);
Tensor<SaveMeanInvStdDataType> save_mean({N, G});
Tensor<SaveMeanInvStdDataType> save_inv_std({N, G});
Tensor<YDataType> host_y(length); Tensor<YDataType> host_y(length);
Tensor<SaveMeanInvStdDataType> host_save_mean({N, G});
Tensor<SaveMeanInvStdDataType> host_save_inv_std({N, G});
std::vector<index_t> strideSaveMeanInvStd = {1};
switch(init_method) switch(init_method)
{ {
...@@ -69,6 +79,9 @@ bool profile_groupnorm_impl(int do_verification, ...@@ -69,6 +79,9 @@ bool profile_groupnorm_impl(int do_verification,
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize()); DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize()); DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize()); DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
DeviceMem save_mean_dev(sizeof(SaveMeanInvStdDataType) * save_mean.mDesc.GetElementSpaceSize());
DeviceMem save_inv_std_dev(sizeof(SaveMeanInvStdDataType) *
save_inv_std.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data()); x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data()); gamma_dev.ToDevice(gamma.mData.data());
...@@ -78,8 +91,8 @@ bool profile_groupnorm_impl(int do_verification, ...@@ -78,8 +91,8 @@ bool profile_groupnorm_impl(int do_verification,
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType, using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
AccDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
PassThrough, PassThrough,
5, 5,
3>; 3>;
...@@ -97,38 +110,70 @@ bool profile_groupnorm_impl(int do_verification, ...@@ -97,38 +110,70 @@ bool profile_groupnorm_impl(int do_verification,
if(do_verification) if(do_verification)
{ {
using ReferenceInstance = ck::tensor_operation::host::ReferenceGroupnorm<XDataType, using ReferenceInstance =
GammaDataType, ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
BetaDataType, GammaDataType,
YDataType, BetaDataType,
AccDataType, YDataType,
PassThrough>; SaveMeanInvStdDataType,
ComputeDataType,
PassThrough>;
ReferenceInstance ref; ReferenceInstance ref;
auto ref_argument = ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, length, 1e-6); auto ref_argument = ref.MakeArgument(
auto ref_invoker = ref.MakeInvoker(); x, gamma, beta, host_y, host_save_mean, host_save_inv_std, PassThrough{}, length, 1e-6);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
} }
int num_kernel = 0; int num_kernel = 0;
auto f_get_argument = [&](auto& inst_ptr) {
if constexpr(SaveMeanInvStd)
return inst_ptr->MakeArgumentPointer(
length,
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
gammaBetaStride,
gammaBetaStride,
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_inv_std.mDesc.GetStrides().begin(),
save_inv_std.mDesc.GetStrides().end()},
reduce_dim,
1e-6,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
save_mean_dev.GetDeviceBuffer(),
save_inv_std_dev.GetDeviceBuffer(),
PassThrough{});
else
return inst_ptr->MakeArgumentPointer(
length,
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
gammaBetaStride,
gammaBetaStride,
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_inv_std.mDesc.GetStrides().begin(),
save_inv_std.mDesc.GetStrides().end()},
reduce_dim,
1e-6,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
nullptr,
nullptr,
PassThrough{});
};
for(auto& inst_ptr : instance_ptrs) for(auto& inst_ptr : instance_ptrs)
{ {
auto argument_ptr = inst_ptr->MakeArgumentPointer( auto argument_ptr = f_get_argument(inst_ptr);
length,
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
gammaBetaStride,
gammaBetaStride,
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
reduce_dim,
1e-6,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
nullptr,
nullptr,
PassThrough{});
if(inst_ptr->IsSupportedArgument(argument_ptr.get())) if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{ {
...@@ -152,6 +197,10 @@ bool profile_groupnorm_impl(int do_verification, ...@@ -152,6 +197,10 @@ bool profile_groupnorm_impl(int do_verification,
beta.mDesc.GetElementSize() * sizeof(BetaDataType) + beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
y.mDesc.GetElementSize() * sizeof(YDataType); y.mDesc.GetElementSize() * sizeof(YDataType);
if constexpr(SaveMeanInvStd)
num_bytes += save_mean.mDesc.GetElementSpaceSize() * sizeof(SaveMeanInvStdDataType) +
save_inv_std.mDesc.GetElementSpaceSize() * sizeof(SaveMeanInvStdDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time; float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel) if(time_kernel)
...@@ -168,9 +217,22 @@ bool profile_groupnorm_impl(int do_verification, ...@@ -168,9 +217,22 @@ bool profile_groupnorm_impl(int do_verification,
if(do_verification) if(do_verification)
{ {
y_dev.FromDevice(y.mData.data()); y_dev.FromDevice(y.mData.data());
bool pass = ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3); bool pass = ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
if constexpr(SaveMeanInvStd)
{
save_mean_dev.FromDevice(save_mean.mData.data());
pass &= ck::utils::check_err(
save_mean.mData, host_save_mean.mData, "Error: Incorrect results", 1e-3, 1e-3);
save_inv_std_dev.FromDevice(save_inv_std.mData.data());
pass &= ck::utils::check_err(save_inv_std.mData,
host_save_inv_std.mData,
"Error: Incorrect results",
1e-3,
1e-3);
}
if(do_log) if(do_log)
{ {
LogRangeAsType<float>(std::cout << "x : ", x.mData, ",") << std::endl; LogRangeAsType<float>(std::cout << "x : ", x.mData, ",") << std::endl;
......
profiler/include/profiler/profile_layernorm_impl.hpp
View file @ bbd498f1
...@@ -21,6 +21,8 @@ template <typename XDataType, ...@@ -21,6 +21,8 @@ template <typename XDataType,
typename BetaDataType, typename BetaDataType,
typename ComputeDataType, typename ComputeDataType,
typename YDataType, typename YDataType,
typename SaveMeanInvStdDataType,
bool SaveMeanInvStd,
index_t Rank> index_t Rank>
bool profile_layernorm_impl(int do_verification, bool profile_layernorm_impl(int do_verification,
int init_method, int init_method,
...@@ -43,13 +45,19 @@ bool profile_layernorm_impl(int do_verification, ...@@ -43,13 +45,19 @@ bool profile_layernorm_impl(int do_verification,
Tensor<GammaDataType> gamma(reduce_length); Tensor<GammaDataType> gamma(reduce_length);
Tensor<BetaDataType> beta(reduce_length); Tensor<BetaDataType> beta(reduce_length);
Tensor<YDataType> y(length); Tensor<YDataType> y(length);
Tensor<SaveMeanInvStdDataType> save_mean({length[0]});
Tensor<SaveMeanInvStdDataType> save_inv_std({length[0]});
Tensor<YDataType> host_y(length); Tensor<YDataType> host_y(length);
Tensor<SaveMeanInvStdDataType> host_save_mean({length[0]});
Tensor<SaveMeanInvStdDataType> host_save_inv_std({length[0]});
std::vector<index_t> strideXY = std::vector<index_t> strideXY =
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()}; std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()};
std::vector<index_t> strideGammaBeta = strideXY; std::vector<index_t> strideGammaBeta = strideXY;
strideGammaBeta[0] = 0; strideGammaBeta[0] = 0;
std::vector<index_t> strideSaveMeanInvStd = {1};
switch(init_method) switch(init_method)
{ {
case 0: case 0:
...@@ -75,6 +83,9 @@ bool profile_layernorm_impl(int do_verification, ...@@ -75,6 +83,9 @@ bool profile_layernorm_impl(int do_verification,
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize()); DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize()); DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize()); DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
DeviceMem save_mean_dev(sizeof(SaveMeanInvStdDataType) * save_mean.mDesc.GetElementSpaceSize());
DeviceMem save_inv_std_dev(sizeof(SaveMeanInvStdDataType) *
save_inv_std.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data()); x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data()); gamma_dev.ToDevice(gamma.mData.data());
...@@ -86,8 +97,8 @@ bool profile_layernorm_impl(int do_verification, ...@@ -86,8 +97,8 @@ bool profile_layernorm_impl(int do_verification,
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType, using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
ComputeDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
PassThrough, PassThrough,
Rank, Rank,
NumReduceDim>; NumReduceDim>;
...@@ -105,40 +116,74 @@ bool profile_layernorm_impl(int do_verification, ...@@ -105,40 +116,74 @@ bool profile_layernorm_impl(int do_verification,
if(do_verification) if(do_verification)
{ {
using ReferenceInstance = ck::tensor_operation::host::ReferenceLayernorm<XDataType, using ReferenceInstance =
GammaDataType, ck::tensor_operation::host::ReferenceLayernorm<XDataType,
BetaDataType, GammaDataType,
YDataType, BetaDataType,
ComputeDataType, YDataType,
PassThrough, SaveMeanInvStdDataType,
Rank, ComputeDataType,
NumReduceDim>; PassThrough,
Rank,
NumReduceDim>;
ReferenceInstance ref; ReferenceInstance ref;
auto ref_argument = auto ref_argument = ref.MakeArgument(x,
ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, length, reduce_dim, 1e-4); gamma,
auto ref_invoker = ref.MakeInvoker(); beta,
host_y,
host_save_mean,
host_save_inv_std,
PassThrough{},
length,
reduce_dim,
1e-4);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
} }
int num_kernel = 0; int num_kernel = 0;
auto f_get_argument = [&](auto& inst_ptr) {
if constexpr(SaveMeanInvStd)
return inst_ptr->MakeArgumentPointer(length,
strideXY,
strideGammaBeta,
strideGammaBeta,
strideXY,
strideSaveMeanInvStd,
strideSaveMeanInvStd,
reduce_dim,
1e-4,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
save_mean_dev.GetDeviceBuffer(),
save_inv_std_dev.GetDeviceBuffer(),
PassThrough{});
else
return inst_ptr->MakeArgumentPointer(length,
strideXY,
strideGammaBeta,
strideGammaBeta,
strideXY,
strideSaveMeanInvStd,
strideSaveMeanInvStd,
reduce_dim,
1e-4,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
nullptr,
nullptr,
PassThrough{});
};
for(auto& inst_ptr : instance_ptrs) for(auto& inst_ptr : instance_ptrs)
{ {
auto argument_ptr = inst_ptr->MakeArgumentPointer(length, auto argument_ptr = f_get_argument(inst_ptr);
strideXY,
strideGammaBeta,
strideGammaBeta,
strideXY,
reduce_dim,
1e-4,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
nullptr,
nullptr,
PassThrough{});
if(inst_ptr->IsSupportedArgument(argument_ptr.get())) if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{ {
...@@ -168,6 +213,10 @@ bool profile_layernorm_impl(int do_verification, ...@@ -168,6 +213,10 @@ bool profile_layernorm_impl(int do_verification,
beta.mDesc.GetElementSize() * sizeof(BetaDataType) + beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
y.mDesc.GetElementSize() * sizeof(YDataType); y.mDesc.GetElementSize() * sizeof(YDataType);
if constexpr(SaveMeanInvStd)
num_bytes += save_mean.mDesc.GetElementSpaceSize() * sizeof(SaveMeanInvStdDataType) +
save_inv_std.mDesc.GetElementSpaceSize() * sizeof(SaveMeanInvStdDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time; float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel) if(time_kernel)
...@@ -184,10 +233,23 @@ bool profile_layernorm_impl(int do_verification, ...@@ -184,10 +233,23 @@ bool profile_layernorm_impl(int do_verification,
if(do_verification) if(do_verification)
{ {
y_dev.FromDevice(y.mData.data()); y_dev.FromDevice(y.mData.data());
bool pass = bool pass =
ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3); ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
if constexpr(SaveMeanInvStd)
{
save_mean_dev.FromDevice(save_mean.mData.data());
pass &= ck::utils::check_err(
save_mean.mData, host_save_mean.mData, "Error: Incorrect results", 1e-3, 1e-3);
save_inv_std_dev.FromDevice(save_inv_std.mData.data());
pass &= ck::utils::check_err(save_inv_std.mData,
host_save_inv_std.mData,
"Error: Incorrect results",
1e-3,
1e-3);
}
if(do_log) if(do_log)
{ {
LogRangeAsType<float>(std::cout << "x : ", x.mData, ",") << std::endl; LogRangeAsType<float>(std::cout << "x : ", x.mData, ",") << std::endl;
......
profiler/src/profile_groupnorm.cpp
View file @ bbd498f1
...@@ -93,12 +93,12 @@ int profile_groupnorm(int argc, char* argv[]) ...@@ -93,12 +93,12 @@ int profile_groupnorm(int argc, char* argv[])
if(data_type == ck::DataTypeEnum::Float) if(data_type == ck::DataTypeEnum::Float)
{ {
ck::profiler::profile_groupnorm_impl<F32, F32, F32, F32, F32>( ck::profiler::profile_groupnorm_impl<F32, F32, F32, F32, F32, F32, false>(
do_verification, init_method, do_log, time_kernel, length); do_verification, init_method, do_log, time_kernel, length);
} }
else if(data_type == ck::DataTypeEnum::Half) else if(data_type == ck::DataTypeEnum::Half)
{ {
ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16>( ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16, F32, false>(
do_verification, init_method, do_log, time_kernel, length); do_verification, init_method, do_log, time_kernel, length);
} }
else else
......
profiler/src/profile_layernorm.cpp
View file @ bbd498f1
...@@ -82,12 +82,12 @@ int profile_layernorm(int argc, char* argv[]) ...@@ -82,12 +82,12 @@ int profile_layernorm(int argc, char* argv[])
if(data_type == ck::DataTypeEnum::Half) if(data_type == ck::DataTypeEnum::Half)
{ {
ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, rank>( ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F32, false, rank>(
do_verification, init_method, do_log, time_kernel, length); do_verification, init_method, do_log, time_kernel, length);
} }
else if(data_type == ck::DataTypeEnum::Float) else if(data_type == ck::DataTypeEnum::Float)
{ {
ck::profiler::profile_layernorm_impl<F32, F32, F32, F32, F32, rank>( ck::profiler::profile_layernorm_impl<F32, F32, F32, F32, F32, F32, false, rank>(
do_verification, init_method, do_log, time_kernel, length); do_verification, init_method, do_log, time_kernel, length);
} }
else else
......
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