Skip to content

GitLab

Commit 0c823497 authored by muozturk's avatar muozturk
Browse files

merge

parents 334cfe1c 68f2b5e7
Hide whitespace changes
Inline Side-by-side
Showing with 1969 additions and 310 deletions
include/ck/utility/math.hpp
View file @ 0c823497
......@@ -150,28 +150,6 @@ __host__ __device__ constexpr T clamp(const T& x, const T& lowerbound, const T&
return min(max(x, lowerbound), upperbound);
}
// disallow implicit type casting
template <typename T>
__device__ T exp(T x);
// TODO: add f16 support using v_exp_f16
template <>
__device__ float exp<float>(float x)
{
return __expf(x);
}
template <>
__device__ double exp<double>(double x)
{
return exp(x);
}
static inline __host__ float exp(float x) { return std::expf(x); }
static inline __host__ double exp(double x) { return std::exp(x); }
// greatest common divisor, aka highest common factor
__host__ __device__ constexpr index_t gcd(index_t x, index_t y)
{
......
include/ck/utility/math_v2.hpp
View file @ 0c823497
......@@ -9,6 +9,7 @@
#include "ck/utility/data_type.hpp"
#include "ck/utility/type.hpp"
#include "ck/utility/type_convert.hpp"
namespace ck {
namespace math {
......@@ -92,14 +93,96 @@ static inline __host__ float sqrt(float x) { return std::sqrt(x); };
static inline __host__ double sqrt(double x) { return std::sqrt(x); };
static inline __host__ half_t tanh(half_t x)
template <typename T>
inline __host__ T tanh(T x)
{
return static_cast<half_t>(std::tanh(static_cast<float>(x)));
return ck::type_convert<T>(std::tanhf(ck::type_convert<float>(x)));
};
static inline __host__ float tanh(float x) { return std::tanh(x); };
template <>
inline __host__ float tanh<float>(float x)
{
return std::tanhf(x);
};
template <>
inline __host__ double tanh<double>(double x)
{
return std::tanh(x);
};
template <typename T>
inline __host__ T exp(T x)
{
return ck::type_convert<T>(std::expf(ck::type_convert<float>(x)));
}
template <>
inline __host__ float exp<float>(float x)
{
return std::expf(x);
}
static inline __host__ double tanh(double x) { return std::tanh(x); };
template <>
inline __host__ double exp<double>(double x)
{
return std::exp(x);
}
template <typename T>
inline __host__ T log(T x)
{
return ck::type_convert<T>(std::logf(ck::type_convert<float>(x)));
}
template <>
inline __host__ float log<float>(float x)
{
return std::logf(x);
}
template <>
inline __host__ double log<double>(double x)
{
return std::log(x);
}
template <typename T>
inline __host__ T pow(T x, T gamma)
{
return ck::type_convert<T>(
std::powf(ck::type_convert<float>(x), ck::type_convert<float>(gamma)));
}
template <>
inline __host__ float pow<float>(float x, float gamma)
{
return std::powf(x, gamma);
}
template <>
inline __host__ double pow<double>(double x, double gamma)
{
return std::pow(x, gamma);
}
template <typename T>
inline __host__ T expm1(T x)
{
return ck::type_convert<T>(std::expm1f(ck::type_convert<float>(x)));
}
template <>
inline __host__ float expm1<float>(float x)
{
return std::expm1f(x);
}
template <>
inline __host__ double expm1<double>(double x)
{
return std::expm1(x);
}
// math functions for the HIP kernel, some are implemented by calling hip builtin functions
......@@ -181,14 +264,107 @@ static inline __device__ float sqrt(float x) { return __builtin_amdgcn_sqrtf(x);
static inline __device__ double sqrt(double x) { return __builtin_amdgcn_sqrt(x); };
static inline __device__ half_t tanh(half_t x)
template <typename T>
inline __device__ T tanh(T x)
{
return ck::type_convert<T>(::tanhf(ck::type_convert<float>(x)));
};
template <>
inline __device__ float tanh<float>(float x)
{
return static_cast<half_t>(::tanhf(static_cast<float>(x)));
return ::tanhf(x);
};
static inline __device__ float tanh(float x) { return ::tanhf(x); };
template <>
inline __device__ double tanh<double>(double x)
{
return ::tanh(x);
};
template <typename T>
inline __device__ T exp(T x)
{
return ck::type_convert<T>(__expf(ck::type_convert<float>(x)));
};
template <>
inline __device__ half_t exp<half_t>(half_t x)
{
return hexp(x);
};
template <>
inline __device__ float exp<float>(float x)
{
return __expf(x);
};
static inline __device__ double tanh(double x) { return ::tanh(x); };
template <>
inline __device__ double exp<double>(double x)
{
return exp(x);
};
template <typename T>
inline __device__ T log(T x)
{
return ck::type_convert<T>(__logf(ck::type_convert<float>(x)));
};
template <>
inline __device__ half_t log<half_t>(half_t x)
{
return hlog(x);
};
template <>
inline __device__ float log<float>(float x)
{
return __logf(x);
};
template <>
inline __device__ double log<double>(double x)
{
return log(x);
};
template <typename T>
inline __device__ T pow(T x, T gamma)
{
return ck::type_convert<T>(powf(ck::type_convert<float>(x), ck::type_convert<float>(gamma)));
};
template <>
inline __device__ float pow<float>(float x, float gamma)
{
return powf(x, gamma);
};
template <>
inline __device__ double pow<double>(double x, double gamma)
{
return pow(x, gamma);
};
template <typename T>
inline __device__ T expm1(T x)
{
return ck::type_convert<T>(expm1f(ck::type_convert<float>(x)));
};
template <>
inline __device__ float expm1<float>(float x)
{
return expm1f(x);
};
template <>
inline __device__ double expm1<double>(double x)
{
return expm1(x);
};
} // namespace math
} // namespace ck
include/ck/utility/statically_indexed_array_multi_index.hpp
View file @ 0c823497
......@@ -5,6 +5,7 @@
#define CK_STATICALLY_INDEXED_ARRAY_MULTI_INDEX_HPP
#include "common_header.hpp"
#include "ck/utility/math_v2.hpp"
namespace ck {
......
include/ck/utility/type_convert.hpp
View file @ 0c823497
......@@ -95,12 +95,13 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_
return type_convert<bhalf_t>(x_fp32);
}
#if defined CK_ENABLE_FP8
// convert fp32 to fp8
template <>
inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
float max_fp8 = 240.0f;
x = x > max_fp8 ? max_fp8 : (x < -max_fp8 ? -max_fp8 : x);
union
{
float fval;
......@@ -139,6 +140,36 @@ inline __host__ __device__ float type_convert<float, f8_t>(f8_t x)
#endif
}
template <>
inline __host__ __device__ float2_t type_convert<float2_t, f8x2_t>(f8x2_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
const auto i16val = bit_cast<uint16_t>(x);
return __builtin_amdgcn_cvt_pk_f32_fp8(i16val, 0);
#else
constexpr bool negative_zero_nan = true;
const auto f8x2_v = vector_type<f8_t, 2>(x);
vector_type<float, 2> f32x2_v;
f32x2_v.template AsType<float>()(Number<0>{}) =
utils::cast_from_f8<f8_t, float, negative_zero_nan>(
f8x2_v.template AsType<f8_t>()[Number<0>{}]);
f32x2_v.template AsType<float>()(Number<1>{}) =
utils::cast_from_f8<f8_t, float, negative_zero_nan>(
f8x2_v.template AsType<f8_t>()[Number<1>{}]);
return f32x2_v.template AsType<float2_t>()[Number<0>{}];
#endif
}
template <>
inline __host__ __device__ half2_t type_convert<half2_t, float2_t>(float2_t x)
{
const vector_type<float, 2> f32x2_v(x);
const auto y = __builtin_amdgcn_cvt_pkrtz(f32x2_v.template AsType<float>()[Number<0>{}],
f32x2_v.template AsType<float>()[Number<1>{}]);
return bit_cast<half2_t>(y);
}
// convert fp16 to fp8
template <>
inline __host__ __device__ f8_t type_convert<f8_t, half_t>(half_t x)
......@@ -169,9 +200,7 @@ inline __host__ __device__ half_t type_convert<half_t, f8_t>(f8_t x)
return utils::cast_from_f8<f8_t, half_t, negative_zero_nan>(x);
#endif
}
#endif
#if defined CK_ENABLE_BF8
// convert fp32 to bf8
template <>
inline __host__ __device__ bf8_t type_convert<bf8_t, float>(float x)
......@@ -245,7 +274,6 @@ inline __host__ __device__ half_t type_convert<half_t, bf8_t>(bf8_t x)
return utils::cast_from_f8<bf8_t, half_t, negative_zero_nan>(x);
#endif
}
#endif
// Declare a template function for bf16 conversion using RTN
template <typename Y, typename X>
......@@ -308,7 +336,6 @@ inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, half_t>(h
template <typename Y, typename X>
__host__ __device__ constexpr Y f8_convert_sr(X x);
#if defined CK_ENABLE_FP8
// convert fp32 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
......@@ -355,9 +382,7 @@ inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
x, rng);
#endif
}
#endif
#if defined CK_ENABLE_BF8
// convert fp32 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
......@@ -405,6 +430,5 @@ inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
x, rng);
#endif
}
#endif
} // namespace ck
library/include/ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp
View file @ 0c823497
......@@ -19,9 +19,7 @@ namespace host {
* \brief Reference implementation for column to image.
*
* Input tensor descriptor has [N * Do * Ho * Wo, Z * Y * X * C] data layout.
* Memory layout is the same.
* Output 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].
*
* \tparam NDimSpatial Number of spatial dimensions.
* \tparam ImageLayout Image Layout.
......@@ -95,18 +93,19 @@ struct ReferenceColumnToImage : public device::BaseOperator
float Run(const Argument& arg)
{
if(!(arg.output_.GetNumOfDimension() == NDimSpatial + 3 &&
arg.input_.GetNumOfDimension() == 2))
arg.input_.GetNumOfDimension() == 3))
{
throw std::runtime_error("wrong! inconsistent dimension");
}
const index_t G = arg.output_.GetLengths()[0];
const index_t N = arg.output_.GetLengths()[1];
const index_t C = arg.output_.GetLengths()[2];
if constexpr(NDimSpatial == 1)
{
const index_t Wo = arg.output_spatial_lengths_[0];
auto func = [&](auto n) {
auto func = [&](auto g, auto n) {
for(index_t wo = 0; wo < Wo; ++wo)
{
index_t row = n * Wo + wo;
......@@ -123,9 +122,10 @@ struct ReferenceColumnToImage : public device::BaseOperator
if(wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.output_.GetLengths()[3])
{
float v_in = ck::type_convert<float>(arg.input_(row, column));
float v_out = ck::type_convert<float>(arg.output_(0, n, c, wi));
arg.output_(0, n, c, wi) =
float v_in =
ck::type_convert<float>(arg.input_(g, row, column));
float v_out = ck::type_convert<float>(arg.output_(g, n, c, wi));
arg.output_(g, n, c, wi) =
ck::type_convert<OutDataType>(v_in + v_out);
}
column++;
......@@ -134,7 +134,7 @@ struct ReferenceColumnToImage : public device::BaseOperator
}
};
make_ParallelTensorFunctor(func, N)(std::thread::hardware_concurrency());
make_ParallelTensorFunctor(func, G, N)(std::thread::hardware_concurrency());
return 0;
}
......@@ -143,7 +143,7 @@ struct ReferenceColumnToImage : public device::BaseOperator
const index_t Ho = arg.output_spatial_lengths_[0];
const index_t Wo = arg.output_spatial_lengths_[1];
auto func = [&](auto n) {
auto func = [&](auto g, auto n) {
for(index_t ho = 0; ho < Ho; ++ho)
{
for(index_t wo = 0; wo < Wo; ++wo)
......@@ -176,10 +176,10 @@ struct ReferenceColumnToImage : public device::BaseOperator
arg.output_.GetLengths()[4])
{
float v_in =
ck::type_convert<float>(arg.input_(row, column));
ck::type_convert<float>(arg.input_(g, row, column));
float v_out = ck::type_convert<float>(
arg.output_(0, n, c, hi, wi));
arg.output_(0, n, c, hi, wi) =
arg.output_(g, n, c, hi, wi));
arg.output_(g, n, c, hi, wi) =
ck::type_convert<OutDataType>(v_in + v_out);
}
column++;
......@@ -190,7 +190,7 @@ struct ReferenceColumnToImage : public device::BaseOperator
}
};
make_ParallelTensorFunctor(func, N)(std::thread::hardware_concurrency());
make_ParallelTensorFunctor(func, G, N)(std::thread::hardware_concurrency());
return 0;
}
......@@ -200,7 +200,7 @@ struct ReferenceColumnToImage : public device::BaseOperator
const index_t Ho = arg.output_spatial_lengths_[1];
const index_t Wo = arg.output_spatial_lengths_[2];
auto func = [&](auto n) {
auto func = [&](auto g, auto n) {
for(index_t d_o = 0; d_o < Do; ++d_o)
{
for(index_t ho = 0; ho < Ho; ++ho)
......@@ -245,10 +245,10 @@ struct ReferenceColumnToImage : public device::BaseOperator
arg.output_.GetLengths()[5])
{
float v_in = ck::type_convert<float>(
arg.input_(row, column));
arg.input_(g, row, column));
float v_out = ck::type_convert<float>(
arg.output_(0, n, c, di, hi, wi));
arg.output_(0, n, c, di, hi, wi) =
arg.output_(g, n, c, di, hi, wi));
arg.output_(g, n, c, di, hi, wi) =
ck::type_convert<OutDataType>(v_in + v_out);
}
column++;
......@@ -261,7 +261,7 @@ struct ReferenceColumnToImage : public device::BaseOperator
}
};
make_ParallelTensorFunctor(func, N)(std::thread::hardware_concurrency());
make_ParallelTensorFunctor(func, G, N)(std::thread::hardware_concurrency());
return 0;
}
......@@ -303,8 +303,9 @@ struct ReferenceColumnToImage : public device::BaseOperator
C * ck::accumulate_n<index_t>(
arg.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
if(!(arg.input_.GetLengths()[0] == static_cast<std::size_t>(NDoHoWo) &&
arg.input_.GetLengths()[1] == static_cast<std::size_t>(CZYX)))
if(!(arg.input_.GetLengths()[0] == static_cast<std::size_t>(G) &&
arg.input_.GetLengths()[1] == static_cast<std::size_t>(NDoHoWo) &&
arg.input_.GetLengths()[2] == static_cast<std::size_t>(CZYX)))
{
return false;
}
......
library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp
View file @ 0c823497
......@@ -23,6 +23,7 @@ template <ck::index_t NumDimM,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ComputeDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
ck::enable_if_t<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
......@@ -69,19 +70,24 @@ struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::Base
{
for(ck::index_t k1 = 0; k1 < K1; ++k1)
{
// Simulate the possible casting when ComputeDataType is different than the
// A/B data types
ComputeDataType v_a_compute_input =
ck::type_convert<ComputeDataType>(arg.a_ms_ks_(m0, m1, k0, k1));
ComputeDataType v_b_compute_input =
ck::type_convert<ComputeDataType>(arg.b_ns_ks_(n0, n1, k0, k1));
AccDataType v_a;
AccDataType v_b;
arg.a_element_op_(
v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
arg.b_element_op_(
v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
arg.a_element_op_(v_a, ck::type_convert<AccDataType>(v_a_compute_input));
arg.b_element_op_(v_b, ck::type_convert<AccDataType>(v_b_compute_input));
v_acc += v_a * v_b;
}
}
arg.c_ms_ns_(m0, m1, n0, n1) = v_acc;
arg.c_ms_ns_(m0, m1, n0, n1) = ck::type_convert<CDataType>(v_acc);
};
make_ParallelTensorFunctor(f_ms_ns,
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
View file @ 0c823497
......@@ -42,6 +42,7 @@ template <ck::index_t NDimSpatial,
typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation,
ck::index_t NumDTensor = 0,
typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
struct ReferenceConvFwd : public device::BaseOperator
{
......@@ -57,10 +58,12 @@ struct ReferenceConvFwd : public device::BaseOperator
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
OutElementwiseOperation out_element_op,
const std::array<Tensor<OutDataType>, NumDTensor>& d_tensors)
: input_{input},
weight_{weight},
output_{output},
d_tensors_{d_tensors},
conv_strides_{conv_filter_strides},
conv_dilations_{conv_filter_dilations},
in_left_pads_{input_left_pads},
......@@ -75,6 +78,8 @@ struct ReferenceConvFwd : public device::BaseOperator
const Tensor<WeiDataType>& weight_;
Tensor<OutDataType>& output_;
const std::array<Tensor<OutDataType>, NumDTensor>& d_tensors_;
std::vector<index_t> conv_strides_;
std::vector<index_t> conv_dilations_;
std::vector<index_t> in_left_pads_;
......@@ -128,11 +133,28 @@ struct ReferenceConvFwd : public device::BaseOperator
}
}
float v_out;
arg.out_element_op_(v_out, v_acc);
arg.output_(g, n, k, wo) = ck::type_convert<OutDataType>(v_out);
OutDataType v_out;
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
if constexpr(NumDTensor == 0)
{
arg.out_element_op_(v_out, v_acc_converted);
}
else if constexpr(NumDTensor == 1)
{
arg.out_element_op_(v_out, v_acc_converted, arg.d_tensors_[0](g, n, k, wo));
}
else if constexpr(NumDTensor == 2)
{
arg.out_element_op_(v_out,
v_acc_converted,
arg.d_tensors_[0](g, n, k, wo),
arg.d_tensors_[1](g, n, k, wo));
}
else
{
throw std::runtime_error("Output ElementOp not supported in reference.");
}
arg.output_(g, n, k, wo) = v_out;
};
make_ParallelTensorFunctor(func,
......@@ -184,11 +206,29 @@ struct ReferenceConvFwd : public device::BaseOperator
}
}
float v_out;
arg.out_element_op_(v_out, v_acc);
arg.output_(g, n, k, ho, wo) = ck::type_convert<OutDataType>(v_out);
OutDataType v_out;
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
if constexpr(NumDTensor == 0)
{
arg.out_element_op_(v_out, v_acc_converted);
}
else if constexpr(NumDTensor == 1)
{
arg.out_element_op_(
v_out, v_acc_converted, arg.d_tensors_[0](g, n, k, ho, wo));
}
else if constexpr(NumDTensor == 2)
{
arg.out_element_op_(v_out,
v_acc_converted,
arg.d_tensors_[0](g, n, k, ho, wo),
arg.d_tensors_[1](g, n, k, ho, wo));
}
else
{
throw std::runtime_error("Output ElementOp not supported in reference.");
}
arg.output_(g, n, k, ho, wo) = v_out;
};
make_ParallelTensorFunctor(func,
......@@ -253,11 +293,29 @@ struct ReferenceConvFwd : public device::BaseOperator
}
}
float v_out;
arg.out_element_op_(v_out, v_acc);
arg.output_(g, n, k, d_o, ho, wo) = ck::type_convert<OutDataType>(v_out);
OutDataType v_out;
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
if constexpr(NumDTensor == 0)
{
arg.out_element_op_(v_out, v_acc_converted);
}
else if constexpr(NumDTensor == 1)
{
arg.out_element_op_(
v_out, v_acc_converted, arg.d_tensors_[0](g, n, k, d_o, ho, wo));
}
else if constexpr(NumDTensor == 2)
{
arg.out_element_op_(v_out,
v_acc_converted,
arg.d_tensors_[0](g, n, k, d_o, ho, wo),
arg.d_tensors_[1](g, n, k, d_o, ho, wo));
}
else
{
throw std::runtime_error("Output ElementOp not supported in reference.");
}
arg.output_(g, n, k, d_o, ho, wo) = v_out;
};
make_ParallelTensorFunctor(func,
......@@ -300,7 +358,8 @@ struct ReferenceConvFwd : public device::BaseOperator
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
OutElementwiseOperation out_element_op,
const std::array<Tensor<OutDataType>, NumDTensor>& d_tensors = {})
{
return Argument{input,
weight,
......@@ -311,7 +370,8 @@ struct ReferenceConvFwd : public device::BaseOperator
input_right_pads,
in_element_op,
wei_element_op,
out_element_op};
out_element_op,
d_tensors};
}
static auto MakeInvoker() { return Invoker{}; }
......
library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
View file @ 0c823497
......@@ -20,8 +20,9 @@ template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
typename AccDataType,
typename AccElementwiseOperation>
typename SaveMeanInvStdDataType,
typename ComputeDataType,
typename YElementwiseOperation>
struct ReferenceGroupnorm : public device::BaseOperator
{
// x = [N, H, W, G, C]
......@@ -35,14 +36,18 @@ struct ReferenceGroupnorm : public device::BaseOperator
const Tensor<GammaDataType>& gamma,
const Tensor<BetaDataType>& beta,
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,
AccDataType epsilon)
ComputeDataType epsilon)
: x_(x),
gamma_(gamma),
beta_(beta),
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),
epsilon_(epsilon)
{
......@@ -52,9 +57,11 @@ struct ReferenceGroupnorm : public device::BaseOperator
const Tensor<XDataType> gamma_;
const Tensor<XDataType> beta_;
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_;
AccDataType epsilon_;
ComputeDataType epsilon_;
};
// Invoker
......@@ -68,8 +75,8 @@ struct ReferenceGroupnorm : public device::BaseOperator
int G = arg.lengths_[3];
int C = arg.lengths_[4];
Tensor<AccDataType> mean({N, G});
Tensor<AccDataType> var({N, G});
Tensor<ComputeDataType> mean({N, G});
Tensor<ComputeDataType> var({N, G});
// Compute mean & var in [H, W, C] by Welford Algorithm
// TODO - parallel for each HWC
......@@ -78,9 +85,9 @@ struct ReferenceGroupnorm : public device::BaseOperator
{
for(int g = 0; g < G; ++g)
{
AccDataType mean_val = type_convert<AccDataType>(0.0f);
AccDataType var_val = type_convert<AccDataType>(0.0f);
int32_t curr_count = 0;
ComputeDataType mean_val = type_convert<ComputeDataType>(0.0f);
ComputeDataType var_val = type_convert<ComputeDataType>(0.0f);
int32_t curr_count = 0;
for(int h = 0; h < H; ++h)
{
......@@ -89,10 +96,11 @@ struct ReferenceGroupnorm : public device::BaseOperator
for(int c = 0; c < C; ++c)
{
curr_count++;
AccDataType x = type_convert<AccDataType>(arg.x_(n, h, w, g, c));
AccDataType delta = x - mean_val;
ComputeDataType x =
type_convert<ComputeDataType>(arg.x_(n, h, w, g, c));
ComputeDataType delta = x - mean_val;
mean_val += delta / curr_count;
AccDataType delta2 = x - mean_val;
ComputeDataType delta2 = x - mean_val;
var_val += delta * delta2;
}
}
......@@ -100,6 +108,12 @@ struct ReferenceGroupnorm : public device::BaseOperator
mean(n, g) = mean_val;
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
{
for(int c = 0; c < C; ++c)
{
AccDataType x = type_convert<AccDataType>(arg.x_(n, h, w, g, c));
AccDataType gamma = type_convert<AccDataType>(arg.gamma_(g, c));
AccDataType beta = type_convert<AccDataType>(arg.beta_(g, c));
AccDataType mean_val = type_convert<AccDataType>(mean(n, g));
AccDataType var_val = type_convert<AccDataType>(var(n, g));
AccDataType y = gamma * (x - mean_val) /
ck::math::sqrt(arg.epsilon_ + var_val) +
beta;
arg.acc_elementwise_op_(y, y);
ComputeDataType x =
type_convert<ComputeDataType>(arg.x_(n, h, w, g, c));
ComputeDataType gamma =
type_convert<ComputeDataType>(arg.gamma_(g, c));
ComputeDataType beta =
type_convert<ComputeDataType>(arg.beta_(g, c));
ComputeDataType mean_val =
type_convert<ComputeDataType>(mean(n, g));
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);
}
}
......@@ -159,11 +177,14 @@ struct ReferenceGroupnorm : public device::BaseOperator
const Tensor<GammaDataType>& gamma,
const Tensor<BetaDataType>& beta,
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,
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{}; }
......
library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp
View file @ 0c823497
......@@ -19,9 +19,7 @@ namespace host {
* \brief Reference implementation for image to column.
*
* 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 [N * Do * Ho * Wo, Z * Y * X * C] data layout.
* Memory layout is the same.
* Output tensor descriptor has [G * N * Do * Ho * Wo, Z * Y * X * C] data layout.
*
* \tparam NDimSpatial Number of spatial dimensions.
* \tparam ImageLayout Image Layout.
......@@ -95,18 +93,19 @@ struct ReferenceImageToColumn : public device::BaseOperator
float Run(const Argument& arg)
{
if(!(arg.input_.GetNumOfDimension() == NDimSpatial + 3 &&
arg.output_.GetNumOfDimension() == 2))
arg.output_.GetNumOfDimension() == 3))
{
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 C = arg.input_.GetLengths()[2];
if constexpr(NDimSpatial == 1)
{
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 column = 0;
......@@ -121,15 +120,15 @@ struct ReferenceImageToColumn : public device::BaseOperator
if(wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
{
InDataType v_in = arg.input_(0, n, c, wi);
arg.output_(row, column) = ck::type_convert<OutDataType>(v_in);
InDataType v_in = arg.input_(g, n, c, wi);
arg.output_(g, row, column) = ck::type_convert<OutDataType>(v_in);
}
column++;
}
}
};
make_ParallelTensorFunctor(func, N, Wo)(std::thread::hardware_concurrency());
make_ParallelTensorFunctor(func, G, N, Wo)(std::thread::hardware_concurrency());
return 0;
}
......@@ -138,7 +137,7 @@ struct ReferenceImageToColumn : public device::BaseOperator
const index_t Ho = arg.output_spatial_lengths_[0];
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 column = 0;
......@@ -162,8 +161,9 @@ struct ReferenceImageToColumn : public device::BaseOperator
wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4])
{
InDataType v_in = arg.input_(0, n, c, hi, wi);
arg.output_(row, column) = ck::type_convert<OutDataType>(v_in);
InDataType v_in = arg.input_(g, n, c, hi, wi);
arg.output_(g, row, column) =
ck::type_convert<OutDataType>(v_in);
}
column++;
}
......@@ -171,7 +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;
}
......@@ -181,7 +181,7 @@ struct ReferenceImageToColumn : public device::BaseOperator
const index_t Ho = arg.output_spatial_lengths_[1];
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 column = 0;
......@@ -213,8 +213,8 @@ struct ReferenceImageToColumn : public device::BaseOperator
ck::type_convert<std::size_t>(wi) <
arg.input_.GetLengths()[5])
{
InDataType v_in = arg.input_(0, n, c, di, hi, wi);
arg.output_(row, column) =
InDataType v_in = arg.input_(g, n, c, di, hi, wi);
arg.output_(g, row, column) =
ck::type_convert<OutDataType>(v_in);
}
column++;
......@@ -224,7 +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());
return 0;
......@@ -267,8 +267,9 @@ struct ReferenceImageToColumn : public device::BaseOperator
C * ck::accumulate_n<index_t>(
arg.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
if(!(arg.output_.GetLengths()[0] == static_cast<std::size_t>(NDoHoWo) &&
arg.output_.GetLengths()[1] == static_cast<std::size_t>(CZYX)))
if(!(arg.output_.GetLengths()[0] == static_cast<std::size_t>(G) &&
arg.output_.GetLengths()[1] == static_cast<std::size_t>(NDoHoWo) &&
arg.output_.GetLengths()[2] == static_cast<std::size_t>(CZYX)))
{
return false;
}
......
library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
View file @ 0c823497
......@@ -20,14 +20,16 @@ template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
typename AccDataType,
typename AccElementwiseOperation,
typename SaveMeanInvStdDataType,
typename ComputeDataType,
typename YElementwiseOperation,
index_t Rank,
index_t NumReduceDim>
struct ReferenceLayernorm : public device::BaseOperator
{
// 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
struct Argument : public device::BaseArgument
......@@ -36,15 +38,19 @@ struct ReferenceLayernorm : public device::BaseOperator
const Tensor<GammaDataType>& gamma_n,
const Tensor<BetaDataType>& beta_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> reduceDims,
AccDataType epsilon)
ComputeDataType epsilon)
: x_m_n_(x_m_n),
gamma_n_(gamma_n),
beta_n_(beta_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),
reduceDims_(reduceDims),
epsilon_(epsilon)
......@@ -55,22 +61,24 @@ struct ReferenceLayernorm : public device::BaseOperator
const Tensor<XDataType> gamma_n_;
const Tensor<XDataType> beta_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> reduceDims_;
AccDataType epsilon_;
ComputeDataType epsilon_;
};
// Invoker
struct Invoker : public device::BaseInvoker
{
float Run(const Argument& arg)
float Run2D(const Argument& arg)
{
int M = arg.lengths_[0];
int N = arg.lengths_[1];
Tensor<AccDataType> mean({M});
Tensor<AccDataType> var({M});
Tensor<ComputeDataType> mean({M});
Tensor<ComputeDataType> var({M});
for(int m = 0; m < M; ++m)
{
......@@ -79,7 +87,7 @@ struct ReferenceLayernorm : public device::BaseOperator
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;
var(m) += x_val * x_val;
}
......@@ -90,22 +98,91 @@ struct ReferenceLayernorm : public device::BaseOperator
for(int m = 0; m < M; ++m)
{
AccDataType divisor =
static_cast<AccDataType>(1) / ck::math::sqrt(var(m) + arg.epsilon_);
ComputeDataType divisor =
static_cast<ComputeDataType>(1) / ck::math::sqrt(var(m) + arg.epsilon_);
for(int n = 0; n < N; ++n)
{
auto x_val = ck::type_convert<AccDataType>(arg.x_m_n_(m, n));
auto y_val = (x_val - mean(m)) * divisor;
y_val = (y_val * arg.gamma_n_(n)) + arg.beta_n_(n);
arg.acc_elementwise_op_(y_val, y_val);
auto x_val = ck::type_convert<ComputeDataType>(arg.x_m_n_(m, n));
auto gamma_val = ck::type_convert<ComputeDataType>(arg.gamma_n_(n));
auto beta_val = ck::type_convert<ComputeDataType>(arg.beta_n_(n));
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.save_mean_m_(m) = ck::type_convert<SaveMeanInvStdDataType>(mean(m));
arg.save_inv_std_m_(m) = ck::type_convert<SaveMeanInvStdDataType>(divisor);
}
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,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
......@@ -123,30 +200,39 @@ struct ReferenceLayernorm : public device::BaseOperator
{
const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
// TODO - support generic layernorm
if(p_arg_->lengths_.size() != 2)
return false;
if(p_arg_->reduceDims_.size() != 1)
return false;
if(p_arg_->lengths_.size() == 2 && p_arg_->reduceDims_.size() == 1 &&
p_arg_->reduceDims_[0] == 1)
return true;
if(p_arg_->reduceDims_[0] != 1)
return false;
else if(p_arg_->lengths_.size() == 4 && p_arg_->reduceDims_.size() == 3 &&
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,
const Tensor<GammaDataType>& gamma_n,
const Tensor<BetaDataType>& beta_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> reduceDims,
AccDataType epsilon)
ComputeDataType epsilon)
{
return Argument{
x_m_n, gamma_n, beta_n, y_m_n, acc_elementwise_op, lengths, reduceDims, epsilon};
return Argument{x_m_n,
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{}; }
......
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View file @ 0c823497
......@@ -20,15 +20,13 @@ using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using I32 = int32_t;
#if defined CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
#if defined CK_ENABLE_BF8
using BF8 = ck::bf8_t;
#endif
using F8 = ck::f8_t;
using BF8 = ck::bf8_t;
using Empty_Tuple = ck::Tuple<>;
using BF16_Tuple = ck::Tuple<BF16>;
using F16_Tuple = ck::Tuple<F16>;
using F16_F16_Tuple = ck::Tuple<F16, F16>;
......
library/include/ck/library/tensor_operation_instance/gpu/contraction/device_contraction_instance.hpp 0 → 100644
View file @ 0c823497
// 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 @ 0c823497
......@@ -17,7 +17,6 @@ namespace tensor_operation {
namespace device {
namespace instance {
#ifdef CK_ENABLE_FP32
// float
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
......@@ -28,7 +27,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn
F32,
PassThrough,
PassThrough,
Bilinear>>>& instances);
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
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
F32,
PassThrough,
PassThrough,
Bilinear>>>& instances);
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
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
F32,
PassThrough,
PassThrough,
Bilinear>>>& instances);
Bilinear,
F32>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
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
F32,
PassThrough,
PassThrough,
Bilinear>>>& instances);
#endif
Bilinear,
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
// double
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
......@@ -78,7 +185,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn
F64,
PassThrough,
PassThrough,
Bilinear>>>& instances);
Bilinear,
F64>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance(
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
F64,
PassThrough,
PassThrough,
Bilinear>>>& instances);
Bilinear,
F64>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance(
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
F64,
PassThrough,
PassThrough,
Bilinear>>>& instances);
Bilinear,
F64>>>& instances);
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance(
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
F64,
PassThrough,
PassThrough,
Bilinear>>>& instances);
#endif
Bilinear,
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
template <index_t NumDimM,
index_t NumDimN,
......@@ -123,7 +395,8 @@ template <index_t NumDimM,
typename ADataType,
typename BDataType,
typename DDataType,
typename EDataType>
typename EDataType,
typename ComputeDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD<
NumDimM,
NumDimN,
......@@ -134,7 +407,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType,
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,
NumDimN,
......@@ -145,45 +419,125 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType,
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()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP32
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)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
op_ptrs);
if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
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
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)
{
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_mnnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance(
op_ptrs);
if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mknn_instance(
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;
}
};
......
library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
View file @ 0c823497
......@@ -17,7 +17,6 @@ namespace tensor_operation {
namespace device {
namespace instance {
#ifdef CK_ENABLE_FP32
// float
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
......@@ -28,7 +27,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instanc
F32,
PassThrough,
PassThrough,
Scale>>>& instances);
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
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
F32,
PassThrough,
PassThrough,
Scale>>>& instances);
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
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
F32,
PassThrough,
PassThrough,
Scale>>>& instances);
Scale,
F32>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
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
F32,
PassThrough,
PassThrough,
Scale>>>& instances);
#endif
Scale,
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
// double
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
......@@ -78,7 +185,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instanc
F64,
PassThrough,
PassThrough,
Scale>>>& instances);
Scale,
F64>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance(
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
F64,
PassThrough,
PassThrough,
Scale>>>& instances);
Scale,
F64>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance(
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
F64,
PassThrough,
PassThrough,
Scale>>>& instances);
Scale,
F64>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance(
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
F64,
PassThrough,
PassThrough,
Scale>>>& instances);
#endif
Scale,
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
template <index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
typename ADataType,
typename BDataType,
typename EDataType>
typename EDataType,
typename ComputeDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD<
NumDimM,
NumDimN,
......@@ -133,7 +406,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType,
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,
NumDimN,
......@@ -144,7 +418,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
EDataType,
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()
{
......@@ -155,34 +430,113 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
op_ptrs);
if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
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
if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> &&
is_same_v<EDataType, double>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance(
op_ptrs);
if constexpr(is_same_v<ComputeDataType, double>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance(
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;
}
};
......
library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp
View file @ 0c823497
......@@ -19,109 +19,214 @@ namespace instance {
using namespace ck::conv_tensor_rearrange_op;
// GNWC/GNHWC/GNDHWC
// Image to Column
// nhwc, 1d
void add_device_image_to_column_nwc_1d_bf16_instances(
// 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_nwc_1d_f16_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_nwc_1d_f32_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_nwc_1d_i8_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);
// nhwc, 2d
void add_device_image_to_column_nhwc_2d_bf16_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_nhwc_2d_f16_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_nhwc_2d_f32_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_nhwc_2d_i8_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);
// nhwc, 3d
void add_device_image_to_column_ndhwc_3d_bf16_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_ndhwc_3d_f16_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_ndhwc_3d_f32_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_ndhwc_3d_i8_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
// nhwc, 1d
void add_device_column_to_image_nwc_1d_bf16_instances(
// 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_nwc_1d_f16_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_nwc_1d_f32_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_nwc_1d_i8_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);
// nhwc, 2d
void add_device_column_to_image_nhwc_2d_bf16_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_nhwc_2d_f16_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_nhwc_2d_f32_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_nhwc_2d_i8_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);
// nhwc, 3d
void add_device_column_to_image_ndhwc_3d_bf16_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_ndhwc_3d_f16_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_ndhwc_3d_f32_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_ndhwc_3d_i8_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,
......@@ -151,60 +256,120 @@ struct DeviceOperationInstanceFactory<
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_image_to_column_nwc_1d_f32_instances(op_ptrs);
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_nwc_1d_f16_instances(op_ptrs);
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_nwc_1d_bf16_instances(op_ptrs);
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_nwc_1d_i8_instances(op_ptrs);
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_nhwc_2d_f32_instances(op_ptrs);
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_nhwc_2d_f16_instances(op_ptrs);
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_nhwc_2d_bf16_instances(op_ptrs);
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_nhwc_2d_i8_instances(op_ptrs);
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_ndhwc_3d_f32_instances(op_ptrs);
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_ndhwc_3d_f16_instances(op_ptrs);
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_ndhwc_3d_bf16_instances(op_ptrs);
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_ndhwc_3d_i8_instances(op_ptrs);
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);
}
}
}
......@@ -214,60 +379,120 @@ struct DeviceOperationInstanceFactory<
{
if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
{
add_device_column_to_image_nwc_1d_f32_instances(op_ptrs);
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_nwc_1d_f16_instances(op_ptrs);
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_nwc_1d_bf16_instances(op_ptrs);
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_nwc_1d_i8_instances(op_ptrs);
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_nhwc_2d_f32_instances(op_ptrs);
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_nhwc_2d_f16_instances(op_ptrs);
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_nhwc_2d_bf16_instances(op_ptrs);
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_nhwc_2d_i8_instances(op_ptrs);
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_ndhwc_3d_f32_instances(op_ptrs);
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_ndhwc_3d_f16_instances(op_ptrs);
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_ndhwc_3d_bf16_instances(op_ptrs);
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_ndhwc_3d_i8_instances(op_ptrs);
add_device_column_to_image_ndhwgc_3d_i8_instances(op_ptrs);
}
}
}
......
library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
View file @ 0c823497
......@@ -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> &&
is_same_v<OutLayout, NWK>)
{
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
......@@ -267,17 +269,23 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
}
#endif
}
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
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> &&
is_same_v<OutDataType, float>)
{
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
#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
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
......@@ -306,14 +314,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
}
#endif
}
else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
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> &&
is_same_v<OutDataType, float>)
{
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
......
library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
View file @ 0c823497
......@@ -98,30 +98,31 @@ struct DeviceOperationInstanceFactory<
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
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> &&
is_same_v<OutDataType, float>)
{
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, 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_c_shuffle_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
}
......
library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
View file @ 0c823497
......@@ -328,7 +328,18 @@ void add_device_gemm_xdl_c_shuffle_f8_f8_f8_mk_kn_mn_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,
typename BLayout,
typename CLayout,
......@@ -548,6 +559,20 @@ struct DeviceOperationInstanceFactory<
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
return op_ptrs;
}
......
library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
View file @ 0c823497
......@@ -98,6 +98,26 @@ void add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);
#endif
template <typename ADataType,
......@@ -105,7 +125,8 @@ template <typename ADataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
typename CLayout,
typename ComputeType>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
BLayout,
......@@ -115,7 +136,8 @@ struct DeviceOperationInstanceFactory<
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>>
ck::tensor_operation::element_wise::PassThrough,
ComputeType>>
{
using DeviceOp = DeviceGemmSplitK<ALayout,
BLayout,
......@@ -125,14 +147,15 @@ struct DeviceOperationInstanceFactory<
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
ck::tensor_operation::element_wise::PassThrough,
ComputeType>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
is_same_v<CDataType, float>)
is_same_v<CDataType, float> && is_same_v<ComputeType, float>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -157,8 +180,8 @@ struct DeviceOperationInstanceFactory<
}
#endif
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t> && is_same_v<ComputeType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -183,8 +206,8 @@ struct DeviceOperationInstanceFactory<
}
#endif
#if(defined(CK_ENABLE_FP16) || defined(CK_ENABLE_FP8))
else if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t> && is_same_v<ComputeType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -207,8 +230,8 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instances(op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, half_t>)
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, half_t> && is_same_v<ComputeType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -231,6 +254,31 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t> && is_same_v<ComputeType, f8_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_comp_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_splitk_f16_f16_f16_comp_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_splitk_f16_f16_f16_comp_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_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances(op_ptrs);
}
}
#endif
return op_ptrs;
}
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp
View file @ 0c823497
......@@ -6,8 +6,6 @@
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
......
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