Skip to content

GitLab

Commit d27e0691 authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'upstream/develop' into merge_upstream_1129 

also fix regression
parents 0a7174ad a2969aa8
Hide whitespace changes
Inline Side-by-side
Showing with 2720 additions and 342 deletions
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
View file @ d27e0691
...@@ -3,12 +3,23 @@ ...@@ -3,12 +3,23 @@
#pragma once #pragma once
#include <iostream> #include <cmath>
#include <cstdlib>
#include <numeric>
#include <type_traits> #include <type_traits>
#include <sstream> #include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp" #include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
...@@ -22,6 +33,7 @@ namespace host { ...@@ -22,6 +33,7 @@ namespace host {
// Supports both GNCHW/NGCHW as well as GNHWC/NHWGC physical layout // Supports both GNCHW/NGCHW as well as GNHWC/NHWGC physical layout
// as long as dimensions in tensor descriptor is in GNCHW order // as long as dimensions in tensor descriptor is in GNCHW order
// //
// @tparam NDimSpatial Number of spatial dimensions.
// @tparam InDataType Input tensor data type. // @tparam InDataType Input tensor data type.
// @tparam WeiDataType Weights tensor data type. // @tparam WeiDataType Weights tensor data type.
// @tparam OutDataType Output tensor data type. // @tparam OutDataType Output tensor data type.
...@@ -29,7 +41,9 @@ namespace host { ...@@ -29,7 +41,9 @@ namespace host {
// operation. // operation.
// @tparam WeiElementwiseOperation Functor for weights tensor elementwise // @tparam WeiElementwiseOperation Functor for weights tensor elementwise
// operation. // operation.
// @tparam NDimSpatial Number of spatial dimensions. // @tparam NumAElementwiseTensor Number of A elementwise tensors.
// @tparam NumBElementwiseTensor Number of B elementwise tensors.
// @tparam NumDElementwiseTensor Number of D elementwise tensors.
// //
// input descriptor in [G, N, C, Do, Ho, Wo] order // input descriptor in [G, N, C, Do, Ho, Wo] order
// weight descriptor in [G, K, C, Z, Y, X] order // weight descriptor in [G, K, C, Z, Y, X] order
...@@ -42,25 +56,35 @@ template <ck::index_t NDimSpatial, ...@@ -42,25 +56,35 @@ template <ck::index_t NDimSpatial,
typename InElementwiseOperation, typename InElementwiseOperation,
typename WeiElementwiseOperation, typename WeiElementwiseOperation,
typename OutElementwiseOperation, typename OutElementwiseOperation,
ck::index_t NumAElementwiseTensor = 0,
ck::index_t NumBElementwiseTensor = 0,
ck::index_t NumDElementwiseTensor = 0,
typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false> typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
struct ReferenceConvFwd : public device::BaseOperator struct ReferenceConvFwd : public device::BaseOperator
{ {
// Argument // Argument
struct Argument : public device::BaseArgument struct Argument : public device::BaseArgument
{ {
Argument(const Tensor<InDataType>& input, Argument(
const Tensor<WeiDataType>& weight, const Tensor<InDataType>& input,
Tensor<OutDataType>& output, const Tensor<WeiDataType>& weight,
std::vector<ck::index_t> conv_filter_strides, Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> input_left_pads, std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_right_pads, std::vector<ck::index_t> input_left_pads,
InElementwiseOperation in_element_op, std::vector<ck::index_t> input_right_pads,
WeiElementwiseOperation wei_element_op, InElementwiseOperation in_element_op,
OutElementwiseOperation out_element_op) WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
const std::array<Tensor<InDataType>, NumAElementwiseTensor>& elementwise_a_tensors,
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors,
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors)
: input_{input}, : input_{input},
weight_{weight}, weight_{weight},
output_{output}, output_{output},
elementwise_a_tensors_{elementwise_a_tensors},
elementwise_b_tensors_{elementwise_b_tensors},
elementwise_d_tensors_{elementwise_d_tensors},
conv_strides_{conv_filter_strides}, conv_strides_{conv_filter_strides},
conv_dilations_{conv_filter_dilations}, conv_dilations_{conv_filter_dilations},
in_left_pads_{input_left_pads}, in_left_pads_{input_left_pads},
...@@ -75,6 +99,10 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -75,6 +99,10 @@ struct ReferenceConvFwd : public device::BaseOperator
const Tensor<WeiDataType>& weight_; const Tensor<WeiDataType>& weight_;
Tensor<OutDataType>& output_; Tensor<OutDataType>& output_;
const std::array<Tensor<InDataType>, NumAElementwiseTensor>& elementwise_a_tensors_;
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors_;
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors_;
std::vector<index_t> conv_strides_; std::vector<index_t> conv_strides_;
std::vector<index_t> conv_dilations_; std::vector<index_t> conv_dilations_;
std::vector<index_t> in_left_pads_; std::vector<index_t> in_left_pads_;
...@@ -114,25 +142,43 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -114,25 +142,43 @@ struct ReferenceConvFwd : public device::BaseOperator
if(wi >= 0 && if(wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3]) ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
{ {
float v_in; InDataType v_in;
float v_wei; WeiDataType v_wei;
arg.in_element_op_( ExecuteElementwiseOp(arg.in_element_op_,
v_in, ck::type_convert<float>(arg.input_(g, n, c, wi))); arg.elementwise_a_tensors_,
Number<NumAElementwiseTensor>{},
arg.wei_element_op_( v_in,
v_wei, ck::type_convert<float>(arg.weight_(g, k, c, x))); arg.input_(g, n, c, wi),
g,
v_acc += v_in * v_wei; n,
c,
wi);
ExecuteElementwiseOp(arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei,
arg.weight_(g, k, c, x),
g,
k,
c,
x);
v_acc +=
ck::type_convert<float>(v_in) * ck::type_convert<float>(v_wei);
} }
} }
} }
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
float v_out; OutDataType& v_out = arg.output_(g, n, k, wo);
ExecuteElementwiseOp(arg.out_element_op_,
arg.out_element_op_(v_out, v_acc); arg.elementwise_d_tensors_,
Number<NumDElementwiseTensor>{},
arg.output_(g, n, k, wo) = ck::type_convert<OutDataType>(v_out); v_out,
v_acc_converted,
g,
n,
k,
wo);
}; };
make_ParallelTensorFunctor(func, make_ParallelTensorFunctor(func,
...@@ -169,26 +215,47 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -169,26 +215,47 @@ struct ReferenceConvFwd : public device::BaseOperator
wi >= 0 && wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4]) ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4])
{ {
float v_in; InDataType v_in;
float v_wei; WeiDataType v_wei;
arg.in_element_op_( ExecuteElementwiseOp(arg.in_element_op_,
v_in, ck::type_convert<float>(arg.input_(g, n, c, hi, wi))); arg.elementwise_a_tensors_,
Number<NumAElementwiseTensor>{},
arg.wei_element_op_( v_in,
v_wei, ck::type_convert<float>(arg.weight_(g, k, c, y, x))); arg.input_(g, n, c, hi, wi),
g,
v_acc += v_in * v_wei; n,
c,
hi,
wi);
ExecuteElementwiseOp(arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei,
arg.weight_(g, k, c, y, x),
g,
k,
c,
y,
x);
v_acc += ck::type_convert<float>(v_in) *
ck::type_convert<float>(v_wei);
} }
} }
} }
} }
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
float v_out; OutDataType& v_out = arg.output_(g, n, k, ho, wo);
ExecuteElementwiseOp(arg.out_element_op_,
arg.out_element_op_(v_out, v_acc); arg.elementwise_d_tensors_,
Number<NumDElementwiseTensor>{},
arg.output_(g, n, k, ho, wo) = ck::type_convert<OutDataType>(v_out); v_out,
v_acc_converted,
g,
n,
k,
ho,
wo);
}; };
make_ParallelTensorFunctor(func, make_ParallelTensorFunctor(func,
...@@ -235,29 +302,51 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -235,29 +302,51 @@ struct ReferenceConvFwd : public device::BaseOperator
ck::type_convert<std::size_t>(wi) < ck::type_convert<std::size_t>(wi) <
arg.input_.GetLengths()[5]) arg.input_.GetLengths()[5])
{ {
float v_in; InDataType v_in;
float v_wei; WeiDataType v_wei;
arg.in_element_op_(v_in, ExecuteElementwiseOp(arg.in_element_op_,
ck::type_convert<float>( arg.elementwise_a_tensors_,
arg.input_(g, n, c, di, hi, wi))); Number<NumAElementwiseTensor>{},
v_in,
arg.wei_element_op_( arg.input_(g, n, c, di, hi, wi),
v_wei, g,
ck::type_convert<float>(arg.weight_(g, k, c, z, y, x))); n,
c,
v_acc += v_in * v_wei; di,
hi,
wi);
ExecuteElementwiseOp(arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei,
arg.weight_(g, k, c, z, y, x),
g,
k,
c,
z,
y,
x);
v_acc += ck::type_convert<float>(v_in) *
ck::type_convert<float>(v_wei);
} }
} }
} }
} }
} }
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
float v_out; OutDataType& v_out = arg.output_(g, n, k, d_o, ho, wo);
ExecuteElementwiseOp(arg.out_element_op_,
arg.out_element_op_(v_out, v_acc); arg.elementwise_d_tensors_,
Number<NumDElementwiseTensor>{},
arg.output_(g, n, k, d_o, ho, wo) = ck::type_convert<OutDataType>(v_out); v_out,
v_acc_converted,
g,
n,
k,
d_o,
ho,
wo);
}; };
make_ParallelTensorFunctor(func, make_ParallelTensorFunctor(func,
...@@ -280,6 +369,36 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -280,6 +369,36 @@ struct ReferenceConvFwd : public device::BaseOperator
} }
}; };
template <typename... Args,
typename ElementwiseOp,
typename ElementwiseTensor,
typename NumTensor,
typename T>
static void ExecuteElementwiseOp(ElementwiseOp& elementwise_op,
ElementwiseTensor& elementwise_tensors,
NumTensor,
T& y,
const T& x,
Args... dims)
{
if constexpr(NumTensor::value == 0)
{
elementwise_op(y, x);
}
else if constexpr(NumTensor::value == 1)
{
elementwise_op(y, x, elementwise_tensors[0](dims...));
}
else if constexpr(NumTensor::value == 2)
{
elementwise_op(y, x, elementwise_tensors[0](dims...), elementwise_tensors[1](dims...));
}
else
{
throw std::runtime_error("ElementOp not supported in reference.");
}
}
static constexpr bool IsValidCompilationParameter() static constexpr bool IsValidCompilationParameter()
{ {
// TODO: properly implement this check // TODO: properly implement this check
...@@ -291,16 +410,20 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -291,16 +410,20 @@ struct ReferenceConvFwd : public device::BaseOperator
return NDimSpatial >= 1 && NDimSpatial <= 3; return NDimSpatial >= 1 && NDimSpatial <= 3;
} }
static auto MakeArgument(const Tensor<InDataType>& input, static auto MakeArgument(
const Tensor<WeiDataType>& weight, const Tensor<InDataType>& input,
Tensor<OutDataType>& output, const Tensor<WeiDataType>& weight,
std::vector<ck::index_t> conv_filter_strides, Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_dilations, std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> input_left_pads, std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_right_pads, std::vector<ck::index_t> input_left_pads,
InElementwiseOperation in_element_op, std::vector<ck::index_t> input_right_pads,
WeiElementwiseOperation wei_element_op, InElementwiseOperation in_element_op,
OutElementwiseOperation out_element_op) WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
const std::array<Tensor<InDataType>, NumAElementwiseTensor>& elementwise_a_tensors = {},
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors = {},
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors = {})
{ {
return Argument{input, return Argument{input,
weight, weight,
...@@ -311,7 +434,10 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -311,7 +434,10 @@ struct ReferenceConvFwd : public device::BaseOperator
input_right_pads, input_right_pads,
in_element_op, in_element_op,
wei_element_op, wei_element_op,
out_element_op}; out_element_op,
elementwise_a_tensors,
elementwise_b_tensors,
elementwise_d_tensors};
} }
static auto MakeInvoker() { return Invoker{}; } static auto MakeInvoker() { return Invoker{}; }
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
View file @ d27e0691
...@@ -20,7 +20,9 @@ template <typename ADataType, ...@@ -20,7 +20,9 @@ template <typename ADataType,
typename AccDataType, typename AccDataType,
typename AElementwiseOperation, typename AElementwiseOperation,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CElementwiseOperation> typename CElementwiseOperation,
typename ComputeTypeA = ADataType,
typename ComputeTypeB = ComputeTypeA>
struct ReferenceGemm : public device::BaseOperator struct ReferenceGemm : public device::BaseOperator
{ {
// Argument // Argument
...@@ -64,8 +66,8 @@ struct ReferenceGemm : public device::BaseOperator ...@@ -64,8 +66,8 @@ struct ReferenceGemm : public device::BaseOperator
for(int k = 0; k < K; ++k) for(int k = 0; k < K; ++k)
{ {
ADataType v_a; ComputeTypeA v_a;
BDataType v_b; ComputeTypeB v_b;
// use PassThrough instead of ConvertBF16RTN for reference calculation // use PassThrough instead of ConvertBF16RTN for reference calculation
if constexpr(is_same_v<AElementwiseOperation, if constexpr(is_same_v<AElementwiseOperation,
......
library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
View file @ d27e0691
...@@ -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_groupnorm_bwd.hpp 0 → 100644
View file @ d27e0691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include <vector>
#include <algorithm>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
namespace ck {
namespace tensor_operation {
namespace host {
template <typename DYDataType,
typename XDataType,
typename GammaDataType,
typename MeanInvStdDataType,
typename DGammaDataType,
typename DBetaDataType,
typename DXDataType,
typename ComputeDataType>
struct ReferenceGroupnormBwd : public device::BaseOperator
{
// Argument
struct Argument : public device::BaseArgument
{
Argument(const Tensor<DYDataType>& dy_nhwgc,
const Tensor<XDataType>& x_nhwgc,
const Tensor<GammaDataType>& gamma_gc,
const Tensor<MeanInvStdDataType>& mean_ng,
const Tensor<MeanInvStdDataType>& inv_std_ng,
Tensor<DGammaDataType>& dgamma_gc,
Tensor<DBetaDataType>& dbeta_gc,
Tensor<DXDataType>& dx_nhwgc,
const std::vector<index_t> lengths)
: dy_nhwgc_(dy_nhwgc),
x_nhwgc_(x_nhwgc),
gamma_gc_(gamma_gc),
mean_ng_(mean_ng),
inv_std_ng_(inv_std_ng),
dgamma_gc_(dgamma_gc),
dbeta_gc_(dbeta_gc),
dx_nhwgc_(dx_nhwgc),
lengths_(lengths)
{
}
const Tensor<DYDataType>& dy_nhwgc_;
const Tensor<XDataType>& x_nhwgc_;
const Tensor<GammaDataType>& gamma_gc_;
const Tensor<MeanInvStdDataType>& mean_ng_;
const Tensor<MeanInvStdDataType>& inv_std_ng_;
Tensor<DGammaDataType>& dgamma_gc_;
Tensor<DBetaDataType>& dbeta_gc_;
Tensor<DXDataType>& dx_nhwgc_;
std::vector<index_t> lengths_;
};
// Invoker
struct Invoker : public device::BaseInvoker
{
float Run(const Argument& arg)
{
int N = arg.lengths_[0];
int H = arg.lengths_[1];
int W = arg.lengths_[2];
int G = arg.lengths_[3];
int C = arg.lengths_[4];
// Calculate dgamma and dbeta
for(int g = 0; g < G; ++g)
for(int c = 0; c < C; ++c)
{
ComputeDataType dgamma = 0;
ComputeDataType dbeta = 0;
for(int n = 0; n < N; ++n)
for(int h = 0; h < H; ++h)
for(int w = 0; w < W; ++w)
{
ComputeDataType dy =
ck::type_convert<ComputeDataType>(arg.dy_nhwgc_(n, h, w, g, c));
ComputeDataType x =
ck::type_convert<ComputeDataType>(arg.x_nhwgc_(n, h, w, g, c));
ComputeDataType mean =
ck::type_convert<ComputeDataType>(arg.mean_ng_(n, g));
ComputeDataType rstd =
ck::type_convert<ComputeDataType>(arg.inv_std_ng_(n, g));
dgamma += dy * rstd * (x - mean);
dbeta += dy;
}
arg.dgamma_gc_(g, c) = ck::type_convert<DGammaDataType>(dgamma);
arg.dbeta_gc_(g, c) = ck::type_convert<DBetaDataType>(dbeta);
}
// Calculate dx
int reduce_size = H * W * C;
for(int n = 0; n < N; ++n)
for(int g = 0; g < G; ++g)
{
ComputeDataType ds = 0;
ComputeDataType db = 0;
ComputeDataType mean = ck::type_convert<ComputeDataType>(arg.mean_ng_(n, g));
ComputeDataType rstd = ck::type_convert<ComputeDataType>(arg.inv_std_ng_(n, g));
for(int h = 0; h < H; ++h)
for(int w = 0; w < W; ++w)
for(int c = 0; c < C; ++c)
{
ComputeDataType dy =
ck::type_convert<ComputeDataType>(arg.dy_nhwgc_(n, h, w, g, c));
ComputeDataType x =
ck::type_convert<ComputeDataType>(arg.x_nhwgc_(n, h, w, g, c));
ComputeDataType gamma =
ck::type_convert<ComputeDataType>(arg.gamma_gc_(g, c));
ds += dy * gamma * x;
db += dy * gamma;
}
for(int h = 0; h < H; ++h)
for(int w = 0; w < W; ++w)
for(int c = 0; c < C; ++c)
{
ComputeDataType dy =
ck::type_convert<ComputeDataType>(arg.dy_nhwgc_(n, h, w, g, c));
ComputeDataType x =
ck::type_convert<ComputeDataType>(arg.x_nhwgc_(n, h, w, g, c));
ComputeDataType gamma =
ck::type_convert<ComputeDataType>(arg.gamma_gc_(g, c));
ComputeDataType b =
(db * mean - ds) * rstd * rstd * rstd / reduce_size;
ComputeDataType c1 = -b * mean - db * rstd / reduce_size;
arg.dx_nhwgc_(n, h, w, g, c) =
ck::type_convert<DXDataType>(dy * gamma * rstd + b * x + c1);
}
}
return 0;
}
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
static auto MakeArgument(const Tensor<DYDataType>& dy_nhwgc,
const Tensor<XDataType>& x_nhwgc,
const Tensor<GammaDataType>& gamma_gc,
const Tensor<MeanInvStdDataType>& mean_ng,
const Tensor<MeanInvStdDataType>& inv_std_ng,
Tensor<DGammaDataType>& dgamma_gc,
Tensor<DBetaDataType>& dbeta_gc,
Tensor<DXDataType>& dx_nhwgc,
const std::vector<index_t> lengths)
{
return Argument{dy_nhwgc,
x_nhwgc,
gamma_gc,
mean_ng,
inv_std_ng,
dgamma_gc,
dbeta_gc,
dx_nhwgc,
lengths};
}
static auto MakeInvoker() { return Invoker{}; }
virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "ReferenceGroupnormBwd"
<< std::endl;
// clang-format on
return str.str();
}
};
} // namespace host
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp
View file @ d27e0691
...@@ -18,16 +18,16 @@ namespace host { ...@@ -18,16 +18,16 @@ namespace host {
/** /**
* \brief Reference implementation for image to column. * \brief Reference implementation for image to column.
* *
* Tensor descriptor has [G, N, C, Di, Hi, Wi] data layout. * Input tensor descriptor has [G, N, C, Di, Hi, Wi] data layout.
* G must be equal to 1. Memory layout is [G, N, Di, Hi, Wi, C]. * Output tensor descriptor has [G * N * Do * Ho * Wo, Z * Y * X * C] data layout.
* *
* \tparam NDimSpatial Number of spatial dimensions. * \tparam NDimSpatial Number of spatial dimensions.
* \tparam InputLayout Input Layout. * \tparam ImageLayout Image Layout.
* \tparam InDataType Input Data Type. * \tparam InDataType Input Data Type.
* \tparam OutDataType Output Data Type. * \tparam OutDataType Output Data Type.
*/ */
template <ck::index_t NDimSpatial, template <ck::index_t NDimSpatial,
typename InputLayout, typename ImageLayout,
typename InDataType, typename InDataType,
typename OutDataType, typename OutDataType,
typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false> typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
...@@ -93,18 +93,19 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -93,18 +93,19 @@ struct ReferenceImageToColumn : public device::BaseOperator
float Run(const Argument& arg) float Run(const Argument& arg)
{ {
if(!(arg.input_.GetNumOfDimension() == NDimSpatial + 3 && if(!(arg.input_.GetNumOfDimension() == NDimSpatial + 3 &&
arg.output_.GetNumOfDimension() == 2)) arg.output_.GetNumOfDimension() == 3))
{ {
throw std::runtime_error("wrong! inconsistent dimension"); throw std::runtime_error("wrong! inconsistent dimension");
} }
const index_t G = arg.input_.GetLengths()[0];
const index_t N = arg.input_.GetLengths()[1]; const index_t N = arg.input_.GetLengths()[1];
const index_t C = arg.input_.GetLengths()[2]; const index_t C = arg.input_.GetLengths()[2];
if constexpr(NDimSpatial == 1) if constexpr(NDimSpatial == 1)
{ {
const index_t Wo = arg.output_spatial_lengths_[0]; const index_t Wo = arg.output_spatial_lengths_[0];
auto func = [&](auto n, auto wo) { auto func = [&](auto g, auto n, auto wo) {
index_t row = n * Wo + wo; index_t row = n * Wo + wo;
index_t column = 0; index_t column = 0;
...@@ -119,15 +120,15 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -119,15 +120,15 @@ struct ReferenceImageToColumn : public device::BaseOperator
if(wi >= 0 && if(wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3]) ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
{ {
InDataType v_in = arg.input_(0, n, c, wi); InDataType v_in = arg.input_(g, n, c, wi);
arg.output_(row, column) = ck::type_convert<OutDataType>(v_in); arg.output_(g, row, column) = ck::type_convert<OutDataType>(v_in);
} }
column++; column++;
} }
} }
}; };
make_ParallelTensorFunctor(func, N, Wo)(std::thread::hardware_concurrency()); make_ParallelTensorFunctor(func, G, N, Wo)(std::thread::hardware_concurrency());
return 0; return 0;
} }
...@@ -136,7 +137,7 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -136,7 +137,7 @@ struct ReferenceImageToColumn : public device::BaseOperator
const index_t Ho = arg.output_spatial_lengths_[0]; const index_t Ho = arg.output_spatial_lengths_[0];
const index_t Wo = arg.output_spatial_lengths_[1]; const index_t Wo = arg.output_spatial_lengths_[1];
auto func = [&](auto n, auto ho, auto wo) { auto func = [&](auto g, auto n, auto ho, auto wo) {
index_t row = n * Ho * Wo + ho * Wo + wo; index_t row = n * Ho * Wo + ho * Wo + wo;
index_t column = 0; index_t column = 0;
...@@ -160,8 +161,9 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -160,8 +161,9 @@ struct ReferenceImageToColumn : public device::BaseOperator
wi >= 0 && wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4]) ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4])
{ {
InDataType v_in = arg.input_(0, n, c, hi, wi); InDataType v_in = arg.input_(g, n, c, hi, wi);
arg.output_(row, column) = ck::type_convert<OutDataType>(v_in); arg.output_(g, row, column) =
ck::type_convert<OutDataType>(v_in);
} }
column++; column++;
} }
...@@ -169,7 +171,7 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -169,7 +171,7 @@ struct ReferenceImageToColumn : public device::BaseOperator
} }
}; };
make_ParallelTensorFunctor(func, N, Ho, Wo)(std::thread::hardware_concurrency()); make_ParallelTensorFunctor(func, G, N, Ho, Wo)(std::thread::hardware_concurrency());
return 0; return 0;
} }
...@@ -179,7 +181,7 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -179,7 +181,7 @@ struct ReferenceImageToColumn : public device::BaseOperator
const index_t Ho = arg.output_spatial_lengths_[1]; const index_t Ho = arg.output_spatial_lengths_[1];
const index_t Wo = arg.output_spatial_lengths_[2]; const index_t Wo = arg.output_spatial_lengths_[2];
auto func = [&](auto n, auto d_o, auto ho, auto wo) { auto func = [&](auto g, auto n, auto d_o, auto ho, auto wo) {
index_t row = n * Do * Ho * Wo + d_o * Ho * Wo + ho * Wo + wo; index_t row = n * Do * Ho * Wo + d_o * Ho * Wo + ho * Wo + wo;
index_t column = 0; index_t column = 0;
...@@ -211,8 +213,8 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -211,8 +213,8 @@ struct ReferenceImageToColumn : public device::BaseOperator
ck::type_convert<std::size_t>(wi) < ck::type_convert<std::size_t>(wi) <
arg.input_.GetLengths()[5]) arg.input_.GetLengths()[5])
{ {
InDataType v_in = arg.input_(0, n, c, di, hi, wi); InDataType v_in = arg.input_(g, n, c, di, hi, wi);
arg.output_(row, column) = arg.output_(g, row, column) =
ck::type_convert<OutDataType>(v_in); ck::type_convert<OutDataType>(v_in);
} }
column++; column++;
...@@ -222,7 +224,7 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -222,7 +224,7 @@ struct ReferenceImageToColumn : public device::BaseOperator
} }
}; };
make_ParallelTensorFunctor(func, N, Do, Ho, Wo)( make_ParallelTensorFunctor(func, G, N, Do, Ho, Wo)(
std::thread::hardware_concurrency()); std::thread::hardware_concurrency());
return 0; return 0;
...@@ -240,8 +242,8 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -240,8 +242,8 @@ struct ReferenceImageToColumn : public device::BaseOperator
{ {
using namespace tensor_layout::convolution; using namespace tensor_layout::convolution;
if constexpr(!(std::is_same_v<InputLayout, GNWC> || std::is_same_v<InputLayout, GNHWC> || if constexpr(!(std::is_same_v<ImageLayout, GNWC> || std::is_same_v<ImageLayout, GNHWC> ||
std::is_same_v<InputLayout, GNDHWC>)) std::is_same_v<ImageLayout, GNDHWC>))
{ {
return false; return false;
} }
...@@ -265,8 +267,9 @@ struct ReferenceImageToColumn : public device::BaseOperator ...@@ -265,8 +267,9 @@ struct ReferenceImageToColumn : public device::BaseOperator
C * ck::accumulate_n<index_t>( C * ck::accumulate_n<index_t>(
arg.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>()); arg.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
if(!(arg.output_.GetLengths()[0] == static_cast<std::size_t>(NDoHoWo) && if(!(arg.output_.GetLengths()[0] == static_cast<std::size_t>(G) &&
arg.output_.GetLengths()[1] == static_cast<std::size_t>(CZYX))) arg.output_.GetLengths()[1] == static_cast<std::size_t>(NDoHoWo) &&
arg.output_.GetLengths()[2] == static_cast<std::size_t>(CZYX)))
{ {
return false; return false;
} }
......
library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
View file @ d27e0691
...@@ -20,14 +20,16 @@ template <typename XDataType, ...@@ -20,14 +20,16 @@ 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
{ {
// TODO - support generic layernorm // TODO - support generic layernorm
static_assert((Rank == 2 && NumReduceDim == 1), "Only support 2D version so far"); static_assert((Rank == 2 && NumReduceDim == 1) || (Rank == 4 && NumReduceDim == 3),
"Only support 2D & 4D version so far");
// Argument // Argument
struct Argument : public device::BaseArgument struct Argument : public device::BaseArgument
...@@ -36,15 +38,19 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -36,15 +38,19 @@ 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)
...@@ -55,22 +61,24 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -55,22 +61,24 @@ 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
struct Invoker : public device::BaseInvoker struct Invoker : public device::BaseInvoker
{ {
float Run(const Argument& arg) float Run2D(const Argument& arg)
{ {
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 +87,7 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -79,7 +87,7 @@ 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;
} }
...@@ -90,22 +98,91 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -90,22 +98,91 @@ 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;
} }
float Run4D(const Argument& arg)
{
int N = arg.lengths_[0];
int H = arg.lengths_[1];
int W = arg.lengths_[2];
int C = arg.lengths_[3];
Tensor<ComputeDataType> mean({N});
Tensor<ComputeDataType> var({N});
int reduce_length = H * W * C;
for(int n = 0; n < N; ++n)
{
mean(n) = 0;
var(n) = 0;
for(int h = 0; h < H; ++h)
for(int w = 0; w < W; ++w)
for(int c = 0; c < C; ++c)
{
auto x_val = ck::type_convert<ComputeDataType>(arg.x_m_n_(n, h, w, c));
mean(n) += x_val;
var(n) += x_val * x_val;
}
mean(n) = mean(n) / reduce_length;
var(n) = (var(n) / reduce_length) - (mean(n) * mean(n));
}
for(int n = 0; n < N; ++n)
{
ComputeDataType divisor =
static_cast<ComputeDataType>(1) / ck::math::sqrt(var(n) + arg.epsilon_);
for(int h = 0; h < H; ++h)
for(int w = 0; w < W; ++w)
for(int c = 0; c < C; ++c)
{
auto x_val = ck::type_convert<ComputeDataType>(arg.x_m_n_(n, h, w, c));
auto gamma_val =
ck::type_convert<ComputeDataType>(arg.gamma_n_(h, w, c));
auto beta_val = ck::type_convert<ComputeDataType>(arg.beta_n_(h, w, c));
auto y_val = (x_val - mean(n)) * divisor;
y_val = (y_val * gamma_val) + beta_val;
arg.y_elementwise_op_(y_val, y_val);
arg.y_m_n_(n, h, w, c) = ck::type_convert<YDataType>(y_val);
}
arg.save_mean_m_(n) = ck::type_convert<SaveMeanInvStdDataType>(mean(n));
arg.save_inv_std_m_(n) = ck::type_convert<SaveMeanInvStdDataType>(divisor);
}
return 0;
}
float Run(const Argument& arg)
{
if(arg.lengths_.size() == 2)
return Run2D(arg);
else if(arg.lengths_.size() == 4)
return Run4D(arg);
return 0;
}
float Run(const device::BaseArgument* p_arg, float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override const StreamConfig& /* stream_config */ = StreamConfig{}) override
{ {
...@@ -123,30 +200,39 @@ struct ReferenceLayernorm : public device::BaseOperator ...@@ -123,30 +200,39 @@ struct ReferenceLayernorm : public device::BaseOperator
{ {
const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg); const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
// TODO - support generic layernorm if(p_arg_->lengths_.size() == 2 && p_arg_->reduceDims_.size() == 1 &&
if(p_arg_->lengths_.size() != 2) p_arg_->reduceDims_[0] == 1)
return false; return true;
if(p_arg_->reduceDims_.size() != 1)
return false;
if(p_arg_->reduceDims_[0] != 1) else if(p_arg_->lengths_.size() == 4 && p_arg_->reduceDims_.size() == 3 &&
return false; p_arg_->reduceDims_[0] == 1 && p_arg_->reduceDims_[1] == 2 &&
p_arg_->reduceDims_[2] == 3)
return true;
return true; return false;
} }
static auto MakeArgument(const Tensor<XDataType>& x_m_n, static auto MakeArgument(const Tensor<XDataType>& x_m_n,
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/reference_tensor_operation/cpu/reference_layernorm_bwd.hpp 0 → 100644
View file @ d27e0691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include <vector>
#include <algorithm>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
namespace ck {
namespace tensor_operation {
namespace host {
template <typename DYDataType,
typename XDataType,
typename GammaDataType,
typename MeanInvStdDataType,
typename DGammaDataType,
typename DBetaDataType,
typename DXDataType,
typename ComputeDataType>
struct ReferenceLayernormBwd : public device::BaseOperator
{
// Argument
struct Argument : public device::BaseArgument
{
Argument(const Tensor<DYDataType>& dy_m_n,
const Tensor<XDataType>& x_m_n,
const Tensor<GammaDataType>& gamma_n,
const Tensor<MeanInvStdDataType>& mean_m,
const Tensor<MeanInvStdDataType>& inv_std_m,
Tensor<DGammaDataType>& dgamma_n,
Tensor<DBetaDataType>& dbeta_n,
Tensor<DXDataType>& dx_m_n,
const std::vector<index_t> lengths)
: dy_m_n_(dy_m_n),
x_m_n_(x_m_n),
gamma_n_(gamma_n),
mean_m_(mean_m),
inv_std_m_(inv_std_m),
dgamma_n_(dgamma_n),
dbeta_n_(dbeta_n),
dx_m_n_(dx_m_n),
lengths_(lengths)
{
}
const Tensor<DYDataType>& dy_m_n_;
const Tensor<XDataType>& x_m_n_;
const Tensor<GammaDataType>& gamma_n_;
const Tensor<MeanInvStdDataType>& mean_m_;
const Tensor<MeanInvStdDataType>& inv_std_m_;
Tensor<DGammaDataType>& dgamma_n_;
Tensor<DBetaDataType>& dbeta_n_;
Tensor<DXDataType>& dx_m_n_;
std::vector<index_t> lengths_;
};
// Invoker
struct Invoker : public device::BaseInvoker
{
float Run(const Argument& arg)
{
int M = arg.lengths_[0];
int N = arg.lengths_[1];
// Calculate dgamma and dbeta
for(int n = 0; n < N; ++n)
{
ComputeDataType dgamma = 0;
ComputeDataType dbeta = 0;
for(int m = 0; m < M; ++m)
{
ComputeDataType dy = ck::type_convert<ComputeDataType>(arg.dy_m_n_(m, n));
ComputeDataType x = ck::type_convert<ComputeDataType>(arg.x_m_n_(m, n));
ComputeDataType mean = ck::type_convert<ComputeDataType>(arg.mean_m_(m));
ComputeDataType rstd = ck::type_convert<ComputeDataType>(arg.inv_std_m_(m));
dgamma += dy * rstd * (x - mean);
dbeta += dy;
}
arg.dgamma_n_(n) = ck::type_convert<DGammaDataType>(dgamma);
arg.dbeta_n_(n) = ck::type_convert<DBetaDataType>(dbeta);
}
// Calculate dx
for(int m = 0; m < M; ++m)
{
ComputeDataType ds = 0;
ComputeDataType db = 0;
ComputeDataType mean = ck::type_convert<ComputeDataType>(arg.mean_m_(m));
ComputeDataType rstd = ck::type_convert<ComputeDataType>(arg.inv_std_m_(m));
for(int n = 0; n < N; ++n)
{
ComputeDataType dy = ck::type_convert<ComputeDataType>(arg.dy_m_n_(m, n));
ComputeDataType x = ck::type_convert<ComputeDataType>(arg.x_m_n_(m, n));
ComputeDataType gamma = ck::type_convert<ComputeDataType>(arg.gamma_n_(n));
ds += dy * gamma * x;
db += dy * gamma;
}
for(int n = 0; n < N; ++n)
{
ComputeDataType dy = ck::type_convert<ComputeDataType>(arg.dy_m_n_(m, n));
ComputeDataType x = ck::type_convert<ComputeDataType>(arg.x_m_n_(m, n));
ComputeDataType gamma = ck::type_convert<ComputeDataType>(arg.gamma_n_(n));
ComputeDataType b = (db * mean - ds) * rstd * rstd * rstd / N;
ComputeDataType c = -b * mean - db * rstd / N;
arg.dx_m_n_(m, n) = ck::type_convert<DXDataType>(dy * gamma * rstd + b * x + c);
}
}
return 0;
}
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
static auto MakeArgument(const Tensor<DYDataType>& dy_m_n,
const Tensor<XDataType>& x_m_n,
const Tensor<GammaDataType>& gamma_n,
const Tensor<MeanInvStdDataType>& mean_m,
const Tensor<MeanInvStdDataType>& inv_std_m,
Tensor<DGammaDataType>& dgamma_n,
Tensor<DBetaDataType>& dbeta_n,
Tensor<DXDataType>& dx_m_n,
const std::vector<index_t> lengths)
{
return Argument{
dy_m_n, x_m_n, gamma_n, mean_m, inv_std_m, dgamma_n, dbeta_n, dx_m_n, lengths};
}
static auto MakeInvoker() { return Invoker{}; }
virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "ReferenceLayernormBwd"
<< std::endl;
// clang-format on
return str.str();
}
};
} // namespace host
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View file @ d27e0691
...@@ -17,13 +17,16 @@ namespace instance { ...@@ -17,13 +17,16 @@ namespace instance {
using F64 = double; using F64 = double;
using F32 = float; using F32 = float;
using F16 = ck::half_t; using F16 = ck::half_t;
using F8 = ck::f8_t;
using BF16 = ck::bhalf_t; using BF16 = ck::bhalf_t;
using I8 = int8_t; using I8 = int8_t;
using I32 = int32_t; using I32 = int32_t;
using F8 = ck::f8_t;
using BF8 = ck::bf8_t;
using Empty_Tuple = ck::Tuple<>; using Empty_Tuple = ck::Tuple<>;
using BF16_Tuple = ck::Tuple<BF16>;
using F16_Tuple = ck::Tuple<F16>; using F16_Tuple = ck::Tuple<F16>;
using F16_F16_Tuple = ck::Tuple<F16, F16>; using F16_F16_Tuple = ck::Tuple<F16, F16>;
...@@ -31,6 +34,7 @@ using F64_Tuple = ck::Tuple<F64>; ...@@ -31,6 +34,7 @@ using F64_Tuple = ck::Tuple<F64>;
using F32_Tuple = ck::Tuple<F32>; using F32_Tuple = ck::Tuple<F32>;
using I32_Tuple = ck::Tuple<I32>; using I32_Tuple = ck::Tuple<I32>;
using I32_F32_Tuple = ck::Tuple<I32, F32>; using I32_F32_Tuple = ck::Tuple<I32, F32>;
using I8_Tuple = ck::Tuple<I8>;
using F32_F32_Tuple = ck::Tuple<F32, F32>; using F32_F32_Tuple = ck::Tuple<F32, F32>;
......
library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp
View file @ d27e0691
...@@ -16,26 +16,26 @@ namespace tensor_operation { ...@@ -16,26 +16,26 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
// FP16 #ifdef CK_ENABLE_FP16
void add_device_batchnorm_backward_rank_4_3_f16_instances( void add_device_batchnorm_backward_rank_4_3_f16_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchNormBwd<F16, F32, F32, F32, F16, F32, F32, PassThrough, 4, 3>>>&); DeviceBatchNormBwd<F16, F32, F32, F32, F16, F32, F32, PassThrough, 4, 3>>>&);
#endif
// FP32 #ifdef CK_ENABLE_FP32
void add_device_batchnorm_backward_rank_4_3_f32_instances( void add_device_batchnorm_backward_rank_4_3_f32_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchNormBwd<F32, F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&); DeviceBatchNormBwd<F32, F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&);
#endif
// BF16 #ifdef CK_ENABLE_BF16
void add_device_batchnorm_backward_rank_4_3_bf16_instances( void add_device_batchnorm_backward_rank_4_3_bf16_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchNormBwd<BF16, F32, F32, F32, BF16, F32, F32, PassThrough, 4, 3>>>&); DeviceBatchNormBwd<BF16, F32, F32, F32, BF16, F32, F32, PassThrough, 4, 3>>>&);
#endif
// FP64 #ifdef CK_ENABLE_FP64
void add_device_batchnorm_backward_rank_4_3_f64_instances( void add_device_batchnorm_backward_rank_4_3_f64_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchNormBwd<F64, F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&); DeviceBatchNormBwd<F64, F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&);
#endif
template <typename XDataType, template <typename XDataType,
typename DxDataType, typename DxDataType,
typename DyDataType, typename DyDataType,
...@@ -72,7 +72,7 @@ struct DeviceOperationInstanceFactory< ...@@ -72,7 +72,7 @@ struct DeviceOperationInstanceFactory<
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<XDataType, F16> && is_same_v<DxDataType, F32> && if constexpr(is_same_v<XDataType, F16> && is_same_v<DxDataType, F32> &&
is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> && is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
is_same_v<ScaleDataType, F16> && is_same_v<DscaleDbiasDataType, F32> && is_same_v<ScaleDataType, F16> && is_same_v<DscaleDbiasDataType, F32> &&
...@@ -83,37 +83,43 @@ struct DeviceOperationInstanceFactory< ...@@ -83,37 +83,43 @@ struct DeviceOperationInstanceFactory<
add_device_batchnorm_backward_rank_4_3_f16_instances(op_ptrs); add_device_batchnorm_backward_rank_4_3_f16_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<XDataType, F32> && is_same_v<DxDataType, F32> && #endif
is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> && #ifdef CK_ENABLE_FP32
is_same_v<ScaleDataType, F32> && is_same_v<DscaleDbiasDataType, F32> && if constexpr(is_same_v<XDataType, F32> && is_same_v<DxDataType, F32> &&
is_same_v<MeanVarDataType, F32>) is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
is_same_v<ScaleDataType, F32> && is_same_v<DscaleDbiasDataType, F32> &&
is_same_v<MeanVarDataType, F32>)
{ {
if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>) if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
{ {
add_device_batchnorm_backward_rank_4_3_f32_instances(op_ptrs); add_device_batchnorm_backward_rank_4_3_f32_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<XDataType, BF16> && is_same_v<DxDataType, F32> && #endif
is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> && #ifdef CK_ENABLE_BF16
is_same_v<ScaleDataType, BF16> && is_same_v<DscaleDbiasDataType, F32> && if constexpr(is_same_v<XDataType, BF16> && is_same_v<DxDataType, F32> &&
is_same_v<MeanVarDataType, F32>) is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
is_same_v<ScaleDataType, BF16> && is_same_v<DscaleDbiasDataType, F32> &&
is_same_v<MeanVarDataType, F32>)
{ {
if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>) if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
{ {
add_device_batchnorm_backward_rank_4_3_bf16_instances(op_ptrs); add_device_batchnorm_backward_rank_4_3_bf16_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<XDataType, F64> && is_same_v<DxDataType, F64> && #endif
is_same_v<DyDataType, F64> && is_same_v<AccDataType, F64> && #ifdef CK_ENABLE_FP64
is_same_v<ScaleDataType, F64> && is_same_v<DscaleDbiasDataType, F64> && if constexpr(is_same_v<XDataType, F64> && is_same_v<DxDataType, F64> &&
is_same_v<MeanVarDataType, F64>) is_same_v<DyDataType, F64> && is_same_v<AccDataType, F64> &&
is_same_v<ScaleDataType, F64> && is_same_v<DscaleDbiasDataType, F64> &&
is_same_v<MeanVarDataType, F64>)
{ {
if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>) if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
{ {
add_device_batchnorm_backward_rank_4_3_f64_instances(op_ptrs); add_device_batchnorm_backward_rank_4_3_f64_instances(op_ptrs);
} }
} }
#endif
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp
View file @ d27e0691
...@@ -16,26 +16,26 @@ namespace tensor_operation { ...@@ -16,26 +16,26 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
// FP16 #ifdef CK_ENABLE_FP16
void add_device_batchnorm_forward_rank_4_3_f16_instances( void add_device_batchnorm_forward_rank_4_3_f16_instances(
std::vector< std::vector<
std::unique_ptr<DeviceBatchNormFwd<F16, F16, F32, F16, F16, F32, PassThrough, 4, 3>>>&); std::unique_ptr<DeviceBatchNormFwd<F16, F16, F32, F16, F16, F32, PassThrough, 4, 3>>>&);
#endif
// FP32 #ifdef CK_ENABLE_FP32
void add_device_batchnorm_forward_rank_4_3_f32_instances( void add_device_batchnorm_forward_rank_4_3_f32_instances(
std::vector< std::vector<
std::unique_ptr<DeviceBatchNormFwd<F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&); std::unique_ptr<DeviceBatchNormFwd<F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&);
#endif
// BF16 #ifdef CK_ENABLE_BF16
void add_device_batchnorm_forward_rank_4_3_bf16_instances( void add_device_batchnorm_forward_rank_4_3_bf16_instances(
std::vector< std::vector<
std::unique_ptr<DeviceBatchNormFwd<BF16, BF16, F32, BF16, BF16, F32, PassThrough, 4, 3>>>&); std::unique_ptr<DeviceBatchNormFwd<BF16, BF16, F32, BF16, BF16, F32, PassThrough, 4, 3>>>&);
#endif
// FP64 #ifdef CK_ENABLE_FP64
void add_device_batchnorm_forward_rank_4_3_f64_instances( void add_device_batchnorm_forward_rank_4_3_f64_instances(
std::vector< std::vector<
std::unique_ptr<DeviceBatchNormFwd<F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&); std::unique_ptr<DeviceBatchNormFwd<F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&);
#endif
template <typename XDataType, template <typename XDataType,
typename YDataType, typename YDataType,
typename AccDataType, typename AccDataType,
...@@ -69,7 +69,7 @@ struct DeviceOperationInstanceFactory< ...@@ -69,7 +69,7 @@ struct DeviceOperationInstanceFactory<
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<XDataType, F16> && is_same_v<YDataType, F16> && if constexpr(is_same_v<XDataType, F16> && is_same_v<YDataType, F16> &&
is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, F16> && is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, F16> &&
is_same_v<BiasDataType, F16> && is_same_v<MeanVarDataType, F32>) is_same_v<BiasDataType, F16> && is_same_v<MeanVarDataType, F32>)
...@@ -79,34 +79,40 @@ struct DeviceOperationInstanceFactory< ...@@ -79,34 +79,40 @@ struct DeviceOperationInstanceFactory<
add_device_batchnorm_forward_rank_4_3_f16_instances(op_ptrs); add_device_batchnorm_forward_rank_4_3_f16_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<XDataType, F32> && is_same_v<YDataType, F32> && #endif
is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, F32> && #ifdef CK_ENABLE_FP32
is_same_v<BiasDataType, F32> && is_same_v<MeanVarDataType, F32>) if constexpr(is_same_v<XDataType, F32> && is_same_v<YDataType, F32> &&
is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, F32> &&
is_same_v<BiasDataType, F32> && is_same_v<MeanVarDataType, F32>)
{ {
if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>) if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
{ {
add_device_batchnorm_forward_rank_4_3_f32_instances(op_ptrs); add_device_batchnorm_forward_rank_4_3_f32_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<XDataType, BF16> && is_same_v<YDataType, BF16> && #endif
is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, BF16> && #ifdef CK_ENABLE_BF16
is_same_v<BiasDataType, BF16> && is_same_v<MeanVarDataType, F32>) if constexpr(is_same_v<XDataType, BF16> && is_same_v<YDataType, BF16> &&
is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, BF16> &&
is_same_v<BiasDataType, BF16> && is_same_v<MeanVarDataType, F32>)
{ {
if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>) if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
{ {
add_device_batchnorm_forward_rank_4_3_bf16_instances(op_ptrs); add_device_batchnorm_forward_rank_4_3_bf16_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<XDataType, F64> && is_same_v<YDataType, F64> && #endif
is_same_v<AccDataType, F64> && is_same_v<ScaleDataType, F64> && #ifdef CK_ENABLE_FP64
is_same_v<BiasDataType, F64> && is_same_v<MeanVarDataType, F64>) if constexpr(is_same_v<XDataType, F64> && is_same_v<YDataType, F64> &&
is_same_v<AccDataType, F64> && is_same_v<ScaleDataType, F64> &&
is_same_v<BiasDataType, F64> && is_same_v<MeanVarDataType, F64>)
{ {
if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>) if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
{ {
add_device_batchnorm_forward_rank_4_3_f64_instances(op_ptrs); add_device_batchnorm_forward_rank_4_3_f64_instances(op_ptrs);
} }
} }
#endif
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp
View file @ d27e0691
...@@ -16,38 +16,38 @@ namespace tensor_operation { ...@@ -16,38 +16,38 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
// FP16 #ifdef CK_ENABLE_FP16
void add_device_batchnorm_infer_rank_4_f16_instances( void add_device_batchnorm_infer_rank_4_f16_instances(
std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise< std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
ck::Tuple<F16, F32, F32, F16, F16>, ck::Tuple<F16, F32, F32, F16, F16>,
ck::Tuple<F16>, ck::Tuple<F16>,
ck::tensor_operation::element_wise::NormalizeInInfer, ck::tensor_operation::element_wise::NormalizeInInfer,
4>>>&); 4>>>&);
#endif
// FP32 #ifdef CK_ENABLE_FP32
void add_device_batchnorm_infer_rank_4_f32_instances( void add_device_batchnorm_infer_rank_4_f32_instances(
std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise< std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
ck::Tuple<F32, F32, F32, F32, F32>, ck::Tuple<F32, F32, F32, F32, F32>,
ck::Tuple<F32>, ck::Tuple<F32>,
ck::tensor_operation::element_wise::NormalizeInInfer, ck::tensor_operation::element_wise::NormalizeInInfer,
4>>>&); 4>>>&);
#endif
// BF16 #ifdef CK_ENABLE_BF16
void add_device_batchnorm_infer_rank_4_bf16_instances( void add_device_batchnorm_infer_rank_4_bf16_instances(
std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise< std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
ck::Tuple<BF16, F32, F32, BF16, BF16>, ck::Tuple<BF16, F32, F32, BF16, BF16>,
ck::Tuple<BF16>, ck::Tuple<BF16>,
ck::tensor_operation::element_wise::NormalizeInInfer, ck::tensor_operation::element_wise::NormalizeInInfer,
4>>>&); 4>>>&);
#endif
// FP64 #ifdef CK_ENABLE_FP64
void add_device_batchnorm_infer_rank_4_f64_instances( void add_device_batchnorm_infer_rank_4_f64_instances(
std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise< std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
ck::Tuple<F64, F64, F64, F64, F64>, ck::Tuple<F64, F64, F64, F64, F64>,
ck::Tuple<F64>, ck::Tuple<F64>,
ck::tensor_operation::element_wise::NormalizeInInfer, ck::tensor_operation::element_wise::NormalizeInInfer,
4>>>&); 4>>>&);
#endif
template <typename XDataType, template <typename XDataType,
typename YDataType, typename YDataType,
typename ScaleDataType, typename ScaleDataType,
...@@ -69,7 +69,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElemen ...@@ -69,7 +69,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElemen
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<XDataType, F16> && is_same_v<YDataType, F16> && if constexpr(is_same_v<XDataType, F16> && is_same_v<YDataType, F16> &&
is_same_v<ScaleDataType, F16> && is_same_v<BiasDataType, F16> && is_same_v<ScaleDataType, F16> && is_same_v<BiasDataType, F16> &&
is_same_v<MeanVarDataType, F32>) is_same_v<MeanVarDataType, F32>)
...@@ -79,34 +79,40 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElemen ...@@ -79,34 +79,40 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElemen
add_device_batchnorm_infer_rank_4_f16_instances(op_ptrs); add_device_batchnorm_infer_rank_4_f16_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<XDataType, F32> && is_same_v<YDataType, F32> && #endif
is_same_v<ScaleDataType, F32> && is_same_v<BiasDataType, F32> && #ifdef CK_ENABLE_FP32
is_same_v<MeanVarDataType, F32>) if constexpr(is_same_v<XDataType, F32> && is_same_v<YDataType, F32> &&
is_same_v<ScaleDataType, F32> && is_same_v<BiasDataType, F32> &&
is_same_v<MeanVarDataType, F32>)
{ {
if constexpr(Rank == 4) if constexpr(Rank == 4)
{ {
add_device_batchnorm_infer_rank_4_f32_instances(op_ptrs); add_device_batchnorm_infer_rank_4_f32_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<XDataType, BF16> && is_same_v<YDataType, BF16> && #endif
is_same_v<ScaleDataType, BF16> && is_same_v<BiasDataType, BF16> && #ifdef CK_ENABLE_BF16
is_same_v<MeanVarDataType, F32>) if constexpr(is_same_v<XDataType, BF16> && is_same_v<YDataType, BF16> &&
is_same_v<ScaleDataType, BF16> && is_same_v<BiasDataType, BF16> &&
is_same_v<MeanVarDataType, F32>)
{ {
if constexpr(Rank == 4) if constexpr(Rank == 4)
{ {
add_device_batchnorm_infer_rank_4_bf16_instances(op_ptrs); add_device_batchnorm_infer_rank_4_bf16_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<XDataType, F64> && is_same_v<YDataType, F64> && #endif
is_same_v<ScaleDataType, F64> && is_same_v<BiasDataType, F64> && #ifdef CK_ENABLE_FP64
is_same_v<MeanVarDataType, F64>) if constexpr(is_same_v<XDataType, F64> && is_same_v<YDataType, F64> &&
is_same_v<ScaleDataType, F64> && is_same_v<BiasDataType, F64> &&
is_same_v<MeanVarDataType, F64>)
{ {
if constexpr(Rank == 4) if constexpr(Rank == 4)
{ {
add_device_batchnorm_infer_rank_4_f64_instances(op_ptrs); add_device_batchnorm_infer_rank_4_f64_instances(op_ptrs);
} }
} }
#endif
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/contraction/device_contraction_instance.hpp 0 → 100644
View file @ d27e0691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using F64 = double;
using F16_Tuple = ck::Tuple<F16>;
using BF16_Tuple = ck::Tuple<BF16>;
using F32_Tuple = ck::Tuple<F32>;
using F64_Tuple = ck::Tuple<F64>;
using Empty_Tuple = ck::Tuple<>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using Scale = ck::tensor_operation::element_wise::Scale;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_kk_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_kn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 1, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 1, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 1, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 1, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 1, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 1, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 1, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 1, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_mk_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 1, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 1, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 1, 4, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 1, 4, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 1, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 1, 4, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 1, 4, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 1, 4, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_mn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 1, 1, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 1, 1, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 1, 1, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 1, 1, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 1, 1, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 1, 1, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 1, 1, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 1, 1, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_f64_kk_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 64, 64, 16, 2, 2, 16, 16, 4, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 2, 2, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 2, 2, 16, 16, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 32, 16, 2, 2, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 32, 128, 16, 2, 2, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 64, 32, 16, 2, 2, 16, 16, 4, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 64, 32, 64, 16, 2, 2, 16, 16, 2, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 1, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_f64_kn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 2, 1, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 2, 1, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 2, 1, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 2, 1, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 2, 2, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 2, 1, 16, 16, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 2, 2, 16, 16, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_f64_mk_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 1, 2, 16, 16, 4, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 1, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 1, 2, 16, 16, 4, 4, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 1, 2, 16, 16, 4, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 2, 2, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 1, 2, 16, 16, 2, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 2, 2, 16, 16, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>
// clang-format on
>;
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ComputeDataType,
typename AElementwiseOp,
typename BElementwiseOp,
typename CDEElementwiseOp>
using device_contraction_f64_mn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Data|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Type|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 1, 1, 16, 16, 4, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 1, 1, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 128, 64, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 1, 1, 16, 16, 4, 4, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 128, 64, 128, 16, 2, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 8, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 1, 1, 16, 16, 4, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 128, 64, 16, 2, 2, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 1, 1, 16, 16, 2, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 0, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementwiseOp, BElementwiseOp, CDEElementwiseOp, GemmMNKPadding, 1, 256, 64, 128, 16, 2, 2, 16, 16, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 16>, 1, ComputeDataType>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
View file @ d27e0691
...@@ -17,7 +17,6 @@ namespace tensor_operation { ...@@ -17,7 +17,6 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
// float
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2, 2,
...@@ -28,7 +27,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn ...@@ -28,7 +27,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -40,7 +40,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn ...@@ -40,7 +40,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -52,7 +53,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn ...@@ -52,7 +53,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -64,10 +66,115 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn ...@@ -64,10 +66,115 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
#endif F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
F16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
F16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
F16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
F16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
BF16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
BF16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
BF16>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
F32_Tuple,
F32,
PassThrough,
PassThrough,
Bilinear,
BF16>>>& instances);
#endif // CK_ENABLE_FP32
#ifdef CK_ENABLE_FP64 #ifdef CK_ENABLE_FP64
// double
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2, 2,
...@@ -78,7 +185,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn ...@@ -78,7 +185,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F64>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -90,7 +198,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn ...@@ -90,7 +198,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F64>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -102,7 +211,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn ...@@ -102,7 +211,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
F64>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -114,8 +224,170 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn ...@@ -114,8 +224,170 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear,
#endif F64>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
F64_Tuple,
F64,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
F64_Tuple,
F64,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
F64_Tuple,
F64,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
F64_Tuple,
F64,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
#endif // CK_ENABLE_FP64
#ifdef CK_ENABLE_FP16
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
F16_Tuple,
F16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
F16_Tuple,
F16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
F16_Tuple,
F16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
F16_Tuple,
F16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
#endif // CK_ENABLE_FP16
#ifdef CK_ENABLE_BF16
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_knnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mnnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
Bilinear,
F32>>>& instances);
#endif // CK_ENABLE_FP16
// Contraction + Bilinear // Contraction + Bilinear
template <index_t NumDimM, template <index_t NumDimM,
index_t NumDimN, index_t NumDimN,
...@@ -123,7 +395,8 @@ template <index_t NumDimM, ...@@ -123,7 +395,8 @@ template <index_t NumDimM,
typename ADataType, typename ADataType,
typename BDataType, typename BDataType,
typename DDataType, typename DDataType,
typename EDataType> typename EDataType,
typename ComputeDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD< struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD<
NumDimM, NumDimM,
NumDimN, NumDimN,
...@@ -134,7 +407,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -134,7 +407,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType, EDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Bilinear>> ck::tensor_operation::element_wise::Bilinear,
ComputeDataType>>
{ {
using DeviceOp = DeviceContractionMultipleD<NumDimM, using DeviceOp = DeviceContractionMultipleD<NumDimM,
NumDimN, NumDimN,
...@@ -145,45 +419,125 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -145,45 +419,125 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType, EDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Bilinear>; ck::tensor_operation::element_wise::Bilinear,
ComputeDataType>;
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> && if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
is_same_v<DDataType, float> && is_same_v<EDataType, float>) is_same_v<EDataType, float>)
{ {
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2) if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{ {
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance( if constexpr(is_same_v<ComputeDataType, float>)
op_ptrs); {
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, ck::half_t>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mnnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, ck::bhalf_t>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mnnn_instance(
op_ptrs);
}
} }
} }
#endif #endif // CK_ENABLE_FP32
#ifdef CK_ENABLE_FP64 #ifdef CK_ENABLE_FP64
if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> && if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> &&
is_same_v<DDataType, double> && is_same_v<EDataType, double>) is_same_v<EDataType, double>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
if constexpr(is_same_v<ComputeDataType, double>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mnnn_instance(
op_ptrs);
}
}
}
#endif // CK_ENABLE_FP64
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<ADataType, ck::half_t> && is_same_v<BDataType, ck::half_t> &&
is_same_v<EDataType, ck::half_t>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mnnn_instance(
op_ptrs);
}
}
}
#endif // CK_ENABLE_FP16
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, ck::bhalf_t> &&
is_same_v<EDataType, ck::bhalf_t>)
{ {
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2) if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{ {
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance( if constexpr(is_same_v<ComputeDataType, float>)
op_ptrs); {
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_kknn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_knnn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance( add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mknn_instance(
op_ptrs); op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mnnn_instance(
op_ptrs);
}
} }
} }
#endif #endif // CK_ENABLE_BF16
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
View file @ d27e0691
...@@ -17,7 +17,6 @@ namespace tensor_operation { ...@@ -17,7 +17,6 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
// float
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2, 2,
...@@ -28,7 +27,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instanc ...@@ -28,7 +27,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instanc
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -40,7 +40,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instanc ...@@ -40,7 +40,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instanc
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -52,7 +53,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instanc ...@@ -52,7 +53,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instanc
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -64,10 +66,115 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instanc ...@@ -64,10 +66,115 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instanc
F32, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
#endif F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
F16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
F16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
F16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
F16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
BF16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
BF16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
BF16>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
Scale,
BF16>>>& instances);
#endif // CK_ENABLE_FP32
#ifdef CK_ENABLE_FP64 #ifdef CK_ENABLE_FP64
// double
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2, 2,
...@@ -78,7 +185,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instanc ...@@ -78,7 +185,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instanc
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F64>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -90,7 +198,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instanc ...@@ -90,7 +198,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instanc
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F64>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -102,7 +211,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instanc ...@@ -102,7 +211,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instanc
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
F64>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -114,15 +224,178 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instanc ...@@ -114,15 +224,178 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instanc
F64, F64,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale,
#endif F64>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
Empty_Tuple,
F64,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
Empty_Tuple,
F64,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
Empty_Tuple,
F64,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F64,
F64,
Empty_Tuple,
F64,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
#endif // CK_ENABLE_FP64
#ifdef CK_ENABLE_FP16
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
#endif // CK_ENABLE_FP16
#ifdef CK_ENABLE_BF16
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
Scale,
F32>>>& instances);
#endif // CK_ENABLE_FP16
// Contraction + Scale // Contraction + Scale
template <index_t NumDimM, template <index_t NumDimM,
index_t NumDimN, index_t NumDimN,
index_t NumDimK, index_t NumDimK,
typename ADataType, typename ADataType,
typename BDataType, typename BDataType,
typename EDataType> typename EDataType,
typename ComputeDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD< struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD<
NumDimM, NumDimM,
NumDimN, NumDimN,
...@@ -133,7 +406,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -133,7 +406,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType, EDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Scale>> ck::tensor_operation::element_wise::Scale,
ComputeDataType>>
{ {
using DeviceOp = DeviceContractionMultipleD<NumDimM, using DeviceOp = DeviceContractionMultipleD<NumDimM,
NumDimN, NumDimN,
...@@ -144,7 +418,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -144,7 +418,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType, EDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Scale>; ck::tensor_operation::element_wise::Scale,
ComputeDataType>;
static auto GetInstances() static auto GetInstances()
{ {
...@@ -155,34 +430,113 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -155,34 +430,113 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
{ {
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2) if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{ {
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance( if constexpr(is_same_v<ComputeDataType, float>)
op_ptrs); {
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, ck::half_t>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, ck::bhalf_t>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mnn_instance(
op_ptrs);
}
} }
} }
#endif #endif // CK_ENABLE_FP32
#ifdef CK_ENABLE_FP64 #ifdef CK_ENABLE_FP64
if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> && if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> &&
is_same_v<EDataType, double>) is_same_v<EDataType, double>)
{ {
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2) if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{ {
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance( if constexpr(is_same_v<ComputeDataType, double>)
op_ptrs); {
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance( add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance(
op_ptrs); op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance(
op_ptrs);
}
else if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance(
op_ptrs);
}
}
}
#endif // CK_ENABLE_FP64
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<ADataType, ck::half_t> && is_same_v<BDataType, ck::half_t> &&
is_same_v<EDataType, ck::half_t>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mnn_instance(
op_ptrs);
}
}
}
#endif // CK_ENABLE_FP16
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, ck::bhalf_t> &&
is_same_v<EDataType, ck::bhalf_t>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mnn_instance(
op_ptrs);
}
} }
} }
#endif #endif // CK_ENABLE_BF16
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp 0 → 100644
View file @ d27e0691
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using namespace ck::conv_tensor_rearrange_op;
// GNWC/GNHWC/GNDHWC
// Image to Column
// GNWC, 1d
void add_device_image_to_column_gnwc_1d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, BF16, BF16, ImageToColumn>>>&
instances);
void add_device_image_to_column_gnwc_1d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, F16, F16, ImageToColumn>>>&
instances);
void add_device_image_to_column_gnwc_1d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, F32, F32, ImageToColumn>>>&
instances);
void add_device_image_to_column_gnwc_1d_i8_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, int8_t, int8_t, ImageToColumn>>>&
instances);
// GNHWC, 2d
void add_device_image_to_column_gnhwc_2d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, BF16, BF16, ImageToColumn>>>&
instances);
void add_device_image_to_column_gnhwc_2d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, F16, F16, ImageToColumn>>>&
instances);
void add_device_image_to_column_gnhwc_2d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, F32, F32, ImageToColumn>>>&
instances);
void add_device_image_to_column_gnhwc_2d_i8_instances(
std::vector<
std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, int8_t, int8_t, ImageToColumn>>>&
instances);
// GNDHWC, 3d
void add_device_image_to_column_gndhwc_3d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, BF16, BF16, ImageToColumn>>>&
instances);
void add_device_image_to_column_gndhwc_3d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, F16, F16, ImageToColumn>>>&
instances);
void add_device_image_to_column_gndhwc_3d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, F32, F32, ImageToColumn>>>&
instances);
void add_device_image_to_column_gndhwc_3d_i8_instances(
std::vector<
std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, int8_t, int8_t, ImageToColumn>>>&
instances);
// Column to Image
// GNWC, 1d
void add_device_column_to_image_gnwc_1d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, BF16, BF16, ColumnToImage>>>&
instances);
void add_device_column_to_image_gnwc_1d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, F16, F16, ColumnToImage>>>&
instances);
void add_device_column_to_image_gnwc_1d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, F32, F32, ColumnToImage>>>&
instances);
void add_device_column_to_image_gnwc_1d_i8_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, int8_t, int8_t, ColumnToImage>>>&
instances);
// GNHWC, 2d
void add_device_column_to_image_gnhwc_2d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, BF16, BF16, ColumnToImage>>>&
instances);
void add_device_column_to_image_gnhwc_2d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, F16, F16, ColumnToImage>>>&
instances);
void add_device_column_to_image_gnhwc_2d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, F32, F32, ColumnToImage>>>&
instances);
void add_device_column_to_image_gnhwc_2d_i8_instances(
std::vector<
std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, int8_t, int8_t, ColumnToImage>>>&
instances);
// GNDHWC, 3d
void add_device_column_to_image_gndhwc_3d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, BF16, BF16, ColumnToImage>>>&
instances);
void add_device_column_to_image_gndhwc_3d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, F16, F16, ColumnToImage>>>&
instances);
void add_device_column_to_image_gndhwc_3d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, F32, F32, ColumnToImage>>>&
instances);
void add_device_column_to_image_gndhwc_3d_i8_instances(
std::vector<
std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, int8_t, int8_t, ColumnToImage>>>&
instances);
// NWGC/NHWGC/NDHWGC
// Image to Column
// NWGC, 1d
void add_device_image_to_column_nwgc_1d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, NWGC, BF16, BF16, ImageToColumn>>>&
instances);
void add_device_image_to_column_nwgc_1d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, NWGC, F16, F16, ImageToColumn>>>&
instances);
void add_device_image_to_column_nwgc_1d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, NWGC, F32, F32, ImageToColumn>>>&
instances);
void add_device_image_to_column_nwgc_1d_i8_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, NWGC, int8_t, int8_t, ImageToColumn>>>&
instances);
// NHWGC, 2d
void add_device_image_to_column_nhwgc_2d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, NHWGC, BF16, BF16, ImageToColumn>>>&
instances);
void add_device_image_to_column_nhwgc_2d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, NHWGC, F16, F16, ImageToColumn>>>&
instances);
void add_device_image_to_column_nhwgc_2d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, NHWGC, F32, F32, ImageToColumn>>>&
instances);
void add_device_image_to_column_nhwgc_2d_i8_instances(
std::vector<
std::unique_ptr<DeviceConvTensorRearrange<2, NHWGC, int8_t, int8_t, ImageToColumn>>>&
instances);
// NDHWGC, 3d
void add_device_image_to_column_ndhwgc_3d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, NDHWGC, BF16, BF16, ImageToColumn>>>&
instances);
void add_device_image_to_column_ndhwgc_3d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, NDHWGC, F16, F16, ImageToColumn>>>&
instances);
void add_device_image_to_column_ndhwgc_3d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, NDHWGC, F32, F32, ImageToColumn>>>&
instances);
void add_device_image_to_column_ndhwgc_3d_i8_instances(
std::vector<
std::unique_ptr<DeviceConvTensorRearrange<3, NDHWGC, int8_t, int8_t, ImageToColumn>>>&
instances);
// Column to Image
// NWGC, 1d
void add_device_column_to_image_nwgc_1d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, NWGC, BF16, BF16, ColumnToImage>>>&
instances);
void add_device_column_to_image_nwgc_1d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, NWGC, F16, F16, ColumnToImage>>>&
instances);
void add_device_column_to_image_nwgc_1d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, NWGC, F32, F32, ColumnToImage>>>&
instances);
void add_device_column_to_image_nwgc_1d_i8_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, NWGC, int8_t, int8_t, ColumnToImage>>>&
instances);
// NHWGC, 2d
void add_device_column_to_image_nhwgc_2d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, NHWGC, BF16, BF16, ColumnToImage>>>&
instances);
void add_device_column_to_image_nhwgc_2d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, NHWGC, F16, F16, ColumnToImage>>>&
instances);
void add_device_column_to_image_nhwgc_2d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, NHWGC, F32, F32, ColumnToImage>>>&
instances);
void add_device_column_to_image_nhwgc_2d_i8_instances(
std::vector<
std::unique_ptr<DeviceConvTensorRearrange<2, NHWGC, int8_t, int8_t, ColumnToImage>>>&
instances);
// NDHWGC, 3d
void add_device_column_to_image_ndhwgc_3d_bf16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, NDHWGC, BF16, BF16, ColumnToImage>>>&
instances);
void add_device_column_to_image_ndhwgc_3d_f16_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, NDHWGC, F16, F16, ColumnToImage>>>&
instances);
void add_device_column_to_image_ndhwgc_3d_f32_instances(
std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, NDHWGC, F32, F32, ColumnToImage>>>&
instances);
void add_device_column_to_image_ndhwgc_3d_i8_instances(
std::vector<
std::unique_ptr<DeviceConvTensorRearrange<3, NDHWGC, int8_t, int8_t, ColumnToImage>>>&
instances);
template <ck::index_t NumDimSpatial,
typename ImageLayout,
typename InDataType,
typename OutDataType,
typename ConvTensorRearrangeOp>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceConvTensorRearrange<NumDimSpatial,
ImageLayout,
InDataType,
OutDataType,
ConvTensorRearrangeOp>>
{
using DeviceOp = DeviceConvTensorRearrange<NumDimSpatial,
ImageLayout,
InDataType,
OutDataType,
ConvTensorRearrangeOp>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(is_same_v<ConvTensorRearrangeOp, ImageToColumn>)
{
if constexpr(NumDimSpatial == 1 && is_same_v<ImageLayout, GNWC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_image_to_column_gnwc_1d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_image_to_column_gnwc_1d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_image_to_column_gnwc_1d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_image_to_column_gnwc_1d_i8_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 2 && is_same_v<ImageLayout, GNHWC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_image_to_column_gnhwc_2d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_image_to_column_gnhwc_2d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_image_to_column_gnhwc_2d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_image_to_column_gnhwc_2d_i8_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 3 && is_same_v<ImageLayout, GNDHWC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_image_to_column_gndhwc_3d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_image_to_column_gndhwc_3d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_image_to_column_gndhwc_3d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_image_to_column_gndhwc_3d_i8_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 1 && is_same_v<ImageLayout, NWGC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_image_to_column_nwgc_1d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_image_to_column_nwgc_1d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_image_to_column_nwgc_1d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_image_to_column_nwgc_1d_i8_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 2 && is_same_v<ImageLayout, NHWGC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_image_to_column_nhwgc_2d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_image_to_column_nhwgc_2d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_image_to_column_nhwgc_2d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_image_to_column_nhwgc_2d_i8_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 3 && is_same_v<ImageLayout, NDHWGC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_image_to_column_ndhwgc_3d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_image_to_column_ndhwgc_3d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_image_to_column_ndhwgc_3d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_image_to_column_ndhwgc_3d_i8_instances(op_ptrs);
}
}
}
else if constexpr(is_same_v<ConvTensorRearrangeOp, ColumnToImage>)
{
if constexpr(NumDimSpatial == 1 && is_same_v<ImageLayout, GNWC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_column_to_image_gnwc_1d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_column_to_image_gnwc_1d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_column_to_image_gnwc_1d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_column_to_image_gnwc_1d_i8_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 2 && is_same_v<ImageLayout, GNHWC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_column_to_image_gnhwc_2d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_column_to_image_gnhwc_2d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_column_to_image_gnhwc_2d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_column_to_image_gnhwc_2d_i8_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 3 && is_same_v<ImageLayout, GNDHWC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_column_to_image_gndhwc_3d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_column_to_image_gndhwc_3d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_column_to_image_gndhwc_3d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_column_to_image_gndhwc_3d_i8_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 1 && is_same_v<ImageLayout, NWGC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_column_to_image_nwgc_1d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_column_to_image_nwgc_1d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_column_to_image_nwgc_1d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_column_to_image_nwgc_1d_i8_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 2 && is_same_v<ImageLayout, NHWGC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_column_to_image_nhwgc_2d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_column_to_image_nhwgc_2d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_column_to_image_nhwgc_2d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_column_to_image_nhwgc_2d_i8_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 3 && is_same_v<ImageLayout, NDHWGC>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_column_to_image_ndhwgc_3d_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
{
add_device_column_to_image_ndhwgc_3d_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_column_to_image_ndhwgc_3d_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
{
add_device_column_to_image_ndhwgc_3d_i8_instances(op_ptrs);
}
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp 0 → 100644
View file @ d27e0691
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using namespace ck::tensor_layout::convolution;
using namespace ck::conv_tensor_rearrange_op;
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
template <ck::index_t NDimSpatial, typename InLayout>
using device_column_to_image_bf16_instances = std::tuple<
// clang-format off
//#####################| Num| InLayout| InDataType| OutDataType| Block| MPer| KPer| Thread| Scalar|
//#####################| Dim| | | | Size| Block| Block| Cluster| Per|
//#####################| Spatial| | | | | | | Lengths| Vector|
//#####################| | | | | | | | | |
// generic instance
DeviceColumnToImageImpl<NDimSpatial, InLayout, BF16, BF16, 64, 16, 16, S<8, 8>, 1>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, BF16, BF16, 64, 32, 32, S<8, 8>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, BF16, BF16, 64, 64, 64, S<8, 8>, 8>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, BF16, BF16, 128, 32, 64, S<8, 16>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, BF16, BF16, 128, 64, 128, S<8, 16>, 8>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, BF16, BF16, 256, 64, 64, S<16, 16>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, BF16, BF16, 256, 128, 128, S<16, 16>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, BF16, BF16, 256, 128, 128, S<16, 16>, 8>
// clang-format on
>;
template <ck::index_t NDimSpatial, typename InLayout>
using device_column_to_image_f16_instances = std::tuple<
// clang-format off
//#####################| Num| InLayout| InDataType| OutDataType| Block| MPer| KPer| Thread| Scalar|
//#####################| Dim| | | | Size| Block| Block| Cluster| Per|
//#####################| Spatial| | | | | | | Lengths| Vector|
//#####################| | | | | | | | | |
// generic instance
DeviceColumnToImageImpl<NDimSpatial, InLayout, F16, F16, 64, 16, 16, S<8, 8>, 1>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, F16, F16, 64, 32, 32, S<8, 8>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, F16, F16, 64, 64, 64, S<8, 8>, 8>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, F16, F16, 128, 32, 64, S<8, 16>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, F16, F16, 128, 64, 128, S<8, 16>, 8>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, F16, F16, 256, 64, 64, S<16, 16>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, F16, F16, 256, 128, 128, S<16, 16>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, F16, F16, 256, 128, 128, S<16, 16>, 8>
// clang-format on
>;
template <ck::index_t NDimSpatial, typename InLayout>
using device_column_to_image_f32_instances = std::tuple<
// clang-format off
//#####################| Num| InLayout| InDataType| OutDataType| Block| MPer| KPer| Thread| Scalar|
//#####################| Dim| | | | Size| Block| Block| Cluster| Per|
//#####################| Spatial| | | | | | | Lengths| Vector|
//#####################| | | | | | | | | |
// generic instance
DeviceColumnToImageImpl<NDimSpatial, InLayout, F32, F32, 64, 16, 16, S<8, 8>, 1>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, F32, F32, 64, 32, 32, S<8, 8>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, F32, F32, 128, 32, 64, S<8, 16>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, F32, F32, 256, 64, 64, S<16, 16>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, F32, F32, 256, 128, 128, S<16, 16>, 4>
// clang-format on
>;
template <ck::index_t NDimSpatial, typename InLayout>
using device_column_to_image_i8_instances = std::tuple<
// clang-format off
//#####################| Num| InLayout| InDataType| OutDataType| Block| MPer| KPer| Thread| Scalar|
//#####################| Dim| | | | Size| Block| Block| Cluster| Per|
//#####################| Spatial| | | | | | | Lengths| Vector|
//#####################| | | | | | | | | |
// generic instance
DeviceColumnToImageImpl<NDimSpatial, InLayout, int8_t, int8_t, 64, 16, 16, S<8, 8>, 1>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, int8_t, int8_t, 64, 32, 32, S<8, 8>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, int8_t, int8_t, 64, 64, 64, S<8, 8>, 8>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, int8_t, int8_t, 128, 32, 64, S<8, 16>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, int8_t, int8_t, 128, 64, 128, S<8, 16>, 8>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 64, 64, S<16, 16>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 128, 128, S<16, 16>, 4>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 128, 128, S<16, 16>, 8>,
DeviceColumnToImageImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 256, 256, S<16, 16>, 16>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/image_to_column/device_image_to_column_instance.hpp → library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp
View file @ d27e0691
...@@ -13,6 +13,7 @@ namespace device { ...@@ -13,6 +13,7 @@ namespace device {
namespace instance { namespace instance {
using namespace ck::tensor_layout::convolution; using namespace ck::tensor_layout::convolution;
using namespace ck::conv_tensor_rearrange_op;
using BF16 = ck::bhalf_t; using BF16 = ck::bhalf_t;
using F16 = ck::half_t; using F16 = ck::half_t;
...@@ -28,17 +29,12 @@ using device_image_to_column_bf16_instances = std::tuple< ...@@ -28,17 +29,12 @@ using device_image_to_column_bf16_instances = std::tuple<
//#####################| Dim| | | | Size| Block| Block| Cluster| Per| //#####################| Dim| | | | Size| Block| Block| Cluster| Per|
//#####################| Spatial| | | | | | | Lengths| Vector| //#####################| Spatial| | | | | | | Lengths| Vector|
//#####################| | | | | | | | | | //#####################| | | | | | | | | |
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 64, 8, 8, S<8, 8>, 1>, // generic instance
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 64, 16, 16, S<8, 8>, 1>, DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 64, 16, 16, S<8, 8>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 64, 32, 32, S<8, 8>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 64, 32, 32, S<8, 8>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 64, 64, 64, S<8, 8>, 8>, DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 64, 64, 64, S<8, 8>, 8>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 128, 16, 16, S<8, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 128, 64, 64, S<8, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 128, 32, 64, S<8, 16>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 128, 32, 64, S<8, 16>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 128, 64, 128, S<8, 16>, 8>, DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 128, 64, 128, S<8, 16>, 8>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 256, 16, 16, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 256, 64, 64, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 256, 128, 128, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 256, 64, 64, S<16, 16>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 256, 64, 64, S<16, 16>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 256, 128, 128, S<16, 16>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 256, 128, 128, S<16, 16>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 256, 128, 128, S<16, 16>, 8> DeviceImageToColumnImpl<NDimSpatial, InLayout, BF16, BF16, 256, 128, 128, S<16, 16>, 8>
...@@ -52,17 +48,13 @@ using device_image_to_column_f16_instances = std::tuple< ...@@ -52,17 +48,13 @@ using device_image_to_column_f16_instances = std::tuple<
//#####################| Dim| | | | Size| Block| Block| Cluster| Per| //#####################| Dim| | | | Size| Block| Block| Cluster| Per|
//#####################| Spatial| | | | | | | Lengths| Vector| //#####################| Spatial| | | | | | | Lengths| Vector|
//#####################| | | | | | | | | | //#####################| | | | | | | | | |
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 64, 8, 8, S<8, 8>, 1>, // generic instance
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 64, 16, 16, S<8, 8>, 1>, DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 64, 16, 16, S<8, 8>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 64, 32, 32, S<8, 8>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 64, 32, 32, S<8, 8>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 64, 64, 64, S<8, 8>, 8>, DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 64, 64, 64, S<8, 8>, 8>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 128, 16, 16, S<8, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 128, 64, 64, S<8, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 128, 32, 64, S<8, 16>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 128, 32, 64, S<8, 16>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 128, 64, 128, S<8, 16>, 8>, DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 128, 64, 128, S<8, 16>, 8>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 256, 16, 16, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 256, 64, 64, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 256, 128, 128, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 256, 64, 64, S<16, 16>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 256, 64, 64, S<16, 16>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 256, 128, 128, S<16, 16>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 256, 128, 128, S<16, 16>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 256, 128, 128, S<16, 16>, 8> DeviceImageToColumnImpl<NDimSpatial, InLayout, F16, F16, 256, 128, 128, S<16, 16>, 8>
...@@ -76,15 +68,11 @@ using device_image_to_column_f32_instances = std::tuple< ...@@ -76,15 +68,11 @@ using device_image_to_column_f32_instances = std::tuple<
//#####################| Dim| | | | Size| Block| Block| Cluster| Per| //#####################| Dim| | | | Size| Block| Block| Cluster| Per|
//#####################| Spatial| | | | | | | Lengths| Vector| //#####################| Spatial| | | | | | | Lengths| Vector|
//#####################| | | | | | | | | | //#####################| | | | | | | | | |
DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 64, 8, 8, S<8, 8>, 1>, // generic instance
DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 64, 16, 16, S<8, 8>, 1>, DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 64, 16, 16, S<8, 8>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 64, 32, 32, S<8, 8>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 64, 32, 32, S<8, 8>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 128, 16, 16, S<8, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 128, 64, 64, S<8, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 128, 32, 64, S<8, 16>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 128, 32, 64, S<8, 16>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 256, 16, 16, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 256, 64, 64, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 256, 128, 128, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 256, 64, 64, S<16, 16>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 256, 64, 64, S<16, 16>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 256, 128, 128, S<16, 16>, 4> DeviceImageToColumnImpl<NDimSpatial, InLayout, F32, F32, 256, 128, 128, S<16, 16>, 4>
// clang-format on // clang-format on
...@@ -97,17 +85,13 @@ using device_image_to_column_i8_instances = std::tuple< ...@@ -97,17 +85,13 @@ using device_image_to_column_i8_instances = std::tuple<
//#####################| Dim| | | | Size| Block| Block| Cluster| Per| //#####################| Dim| | | | Size| Block| Block| Cluster| Per|
//#####################| Spatial| | | | | | | Lengths| Vector| //#####################| Spatial| | | | | | | Lengths| Vector|
//#####################| | | | | | | | | | //#####################| | | | | | | | | |
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 64, 8, 8, S<8, 8>, 1>, // generic instance
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 64, 16, 16, S<8, 8>, 1>, DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 64, 16, 16, S<8, 8>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 64, 32, 32, S<8, 8>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 64, 32, 32, S<8, 8>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 64, 64, 64, S<8, 8>, 8>, DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 64, 64, 64, S<8, 8>, 8>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 128, 16, 16, S<8, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 128, 64, 64, S<8, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 128, 32, 64, S<8, 16>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 128, 32, 64, S<8, 16>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 128, 64, 128, S<8, 16>, 8>, DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 128, 64, 128, S<8, 16>, 8>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 16, 16, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 64, 64, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 128, 128, S<16, 16>, 1>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 64, 64, S<16, 16>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 64, 64, S<16, 16>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 128, 128, S<16, 16>, 4>, DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 128, 128, S<16, 16>, 4>,
DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 128, 128, S<16, 16>, 8>, DeviceImageToColumnImpl<NDimSpatial, InLayout, int8_t, int8_t, 256, 128, 128, S<16, 16>, 8>,
......
library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
View file @ d27e0691
...@@ -240,11 +240,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw ...@@ -240,11 +240,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, NWC> && is_same_v<WeiLayout, KXC> && if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, NWC> && is_same_v<WeiLayout, KXC> &&
is_same_v<OutLayout, NWK>) is_same_v<OutLayout, NWK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(op_ptrs); add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
...@@ -267,17 +269,23 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw ...@@ -267,17 +269,23 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
} }
#endif #endif
} }
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> && if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>) is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs); add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
#ifdef DL_KERNELS }
add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
#endif #endif
#if defined(DL_KERNELS) && defined(CK_ENABLE_FP32)
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
...@@ -306,14 +314,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw ...@@ -306,14 +314,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
} }
#endif #endif
} }
else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> && if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>) is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(op_ptrs); add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
......
library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
View file @ d27e0691
...@@ -98,30 +98,31 @@ struct DeviceOperationInstanceFactory< ...@@ -98,30 +98,31 @@ struct DeviceOperationInstanceFactory<
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> && if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>) is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs); add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
{ {
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs); add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(op_ptrs); add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
is_same_v<OutDataType, ck::bhalf_t>)
{ {
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs); add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t>)
{ {
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs); add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
} }
......
library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
View file @ d27e0691
...@@ -23,12 +23,17 @@ void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances( ...@@ -23,12 +23,17 @@ void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(
DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances( void add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances( void add_device_gemm_dpp_f16_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
...@@ -38,12 +43,17 @@ void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances( ...@@ -38,12 +43,17 @@ void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(
DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances( void add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances( void add_device_gemm_dpp_f16_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
...@@ -53,12 +63,17 @@ void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances( ...@@ -53,12 +63,17 @@ void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(
DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances( void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances( void add_device_gemm_dpp_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
...@@ -68,12 +83,17 @@ void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances( ...@@ -68,12 +83,17 @@ void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances( void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances( void add_device_gemm_dpp_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
...@@ -207,6 +227,10 @@ void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances( ...@@ -207,6 +227,10 @@ void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
...@@ -269,6 +293,26 @@ void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances( ...@@ -269,6 +293,26 @@ void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif #endif
#ifdef CK_ENABLE_FP64 #ifdef CK_ENABLE_FP64
void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances( void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(
...@@ -292,6 +336,34 @@ void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances( ...@@ -292,6 +336,34 @@ void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(
DeviceGemm<Row, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
#endif #endif
#ifdef CK_ENABLE_FP8
void add_device_gemm_xdl_c_shuffle_f8_f8_f8_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F8, F8, F8, PassThrough, PassThrough, PassThrough>>>& instances);
void add_device_gemm_xdl_c_shuffle_f8_f8_f8_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F8, F8, F8, PassThrough, PassThrough, PassThrough>>>& instances);
void add_device_gemm_xdl_c_shuffle_f8_f8_f8_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F8, F8, F8, PassThrough, PassThrough, PassThrough>>>& instances);
void add_device_gemm_xdl_c_shuffle_f8_f8_f8_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F8, F8, F8, PassThrough, PassThrough, PassThrough>>>& instances);
void add_device_gemm_xdl_c_shuffle_f16_f8_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_f16_f8_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
template <typename ALayout, template <typename ALayout,
typename BLayout, typename BLayout,
typename CLayout, typename CLayout,
...@@ -329,38 +401,46 @@ struct DeviceOperationInstanceFactory< ...@@ -329,38 +401,46 @@ struct DeviceOperationInstanceFactory<
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> && if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs); /// add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instances(
op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs); /// add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instances(
op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs); /// add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instances(
op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs); /// add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instances(
op_ptrs);
} }
} }
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
...@@ -370,45 +450,51 @@ struct DeviceOperationInstanceFactory< ...@@ -370,45 +450,51 @@ struct DeviceOperationInstanceFactory<
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> && if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs); /// add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs);
add_device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances(op_ptrs); add_device_gemm_dpp_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
add_device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs); /// add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs);
add_device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_dpp_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instances(
op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs); /// add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs);
add_device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances(op_ptrs); add_device_gemm_dpp_f16_f16_f16_km_kn_mn_instances(op_ptrs);
add_device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs); /// add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs);
add_device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances(op_ptrs); add_device_gemm_dpp_f16_f16_f16_km_nk_mn_instances(op_ptrs);
add_device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs);
#endif #endif
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs);
} }
...@@ -481,6 +567,46 @@ struct DeviceOperationInstanceFactory< ...@@ -481,6 +567,46 @@ struct DeviceOperationInstanceFactory<
#endif #endif
} }
} }
#endif
#ifdef CK_ENABLE_FP8
else if constexpr(is_same_v<ADataType, ck::f8_t> && is_same_v<BDataType, ck::f8_t> &&
is_same_v<CDataType, ck::f8_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_c_shuffle_f8_f8_f8_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_c_shuffle_f8_f8_f8_mk_nk_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_c_shuffle_f8_f8_f8_km_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_c_shuffle_f8_f8_f8_km_nk_mn_instances(op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, ck::half_t> && is_same_v<BDataType, ck::f8_t> &&
is_same_v<CDataType, ck::half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_c_shuffle_f16_f8_f16_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_c_shuffle_f16_f8_f16_mk_nk_mn_instances(op_ptrs);
}
}
#endif #endif
return op_ptrs; return op_ptrs;
} }
......
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