Skip to content

GitLab

Commit 8dd7156d authored by ltqin's avatar ltqin
Browse files

Merge branch 'mha-train-develop' into attn-train-develop-qloop-mask

parents d5f629e7 b5a3ea2d
Hide whitespace changes
Inline Side-by-side
Showing with 1412 additions and 136 deletions
include/ck/tensor_operation/gpu/device/tensor_layout.hpp
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......
include/ck/tensor_operation/gpu/device/tensor_specialization.hpp
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......
include/ck/tensor_operation/gpu/device/welford_helper.hpp
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......
include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
......@@ -49,6 +50,14 @@ struct Add
y = x0 + x1;
};
template <>
__host__ __device__ constexpr void
operator()<float>(float& y, const float& x0, const bhalf_t& x1) const
{
const float x1_tmp = ck::type_convert<float>(x1);
y = x0 + x1_tmp;
}
template <>
__host__ __device__ constexpr void
operator()<bhalf_t>(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const
......@@ -67,6 +76,30 @@ struct Add
};
};
struct ScaleAdd
{
__host__ __device__ ScaleAdd(float scale) : scale_(scale) {}
template <typename Y, typename X0, typename X1>
__host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
template <>
__host__ __device__ void
operator()<float, float, half_t>(float& y, const float& x0, const half_t& x1) const
{
y = scale_ * x0 + ck::type_convert<float>(x1);
};
template <>
__host__ __device__ void
operator()<float, float, bhalf_t>(float& y, const float& x0, const bhalf_t& x1) const
{
y = scale_ * x0 + ck::type_convert<float>(x1);
};
float scale_;
};
struct Subtract
{
template <typename T>
......@@ -118,6 +151,13 @@ struct Bilinear
template <typename Y, typename X0, typename X1>
__host__ __device__ constexpr void operator()(Y&, const X0&, const X1&) const;
template <>
__host__ __device__ constexpr void
operator()<double, double, double>(double& y, const double& x0, const double& x1) const
{
y = alpha_ * x0 + beta_ * x1;
};
template <>
__host__ __device__ constexpr void
operator()<float, float, float>(float& y, const float& x0, const float& x1) const
......@@ -241,43 +281,42 @@ struct AddHardswish
};
};
// C = A * B
// E = FastGelu(C + D)
struct AddFastGelu
{
// Fast GeLU
// https://paperswithcode.com/method/gelu
// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
__host__ __device__ static constexpr float GetFastGeLU(float x)
{
const float u = 2.f * x * (0.035677f * x * x + 0.797885f);
const float emu = exp(-u);
const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
return x * cdf;
}
template <typename T>
static inline constexpr bool is_valid_param_type_v =
std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>;
template <typename E, typename C, typename D>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
template <>
__host__ __device__ constexpr void
operator()<float, float, float>(float& e, const float& c, const float& d) const
{
static_assert(is_valid_param_type_v<E> && is_valid_param_type_v<C> &&
is_valid_param_type_v<D>);
const float x = c + d;
const float y = GetFastGeLU(type_convert<float>(c) + type_convert<float>(d));
FastGelu{}.template operator()<float, float>(e, x);
}
e = type_convert<E>(y);
template <>
__host__ __device__ constexpr void
operator()<half_t, half_t, half_t>(half_t& e, const half_t& c, const half_t& d) const
{
const half_t x = c + d;
ck::tensor_operation::element_wise::FastGelu{}.template operator()<half_t, half_t>(e, x);
}
template <typename D>
__host__ __device__ constexpr void operator()(float& e, const float& c, const D& d) const
template <>
__host__ __device__ constexpr void
operator()<half_t, float, half_t>(half_t& e, const float& c, const half_t& d) const
{
static_assert(is_valid_param_type_v<D>);
const float x0_f = c + d;
float x1_f = 0;
ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
x0_f);
e = GetFastGeLU(c + type_convert<float>(d));
e = type_convert<half_t>(x1_f);
}
};
......
include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -16,7 +16,7 @@ namespace element_wise {
// Need to ensure compiler will fail if there is no matching candidate, instead of compiler
// siliently do implicit type conversion
//
// Method 1:
// Example:
//
// struct ExampleElementwiseOp
// {
......@@ -30,19 +30,6 @@ namespace element_wise {
// {
// }
// };
//
// Method 2:
//
// template <typename Y, typename X>
// struct ExampleElementwiseOp;
//
// template <>
// struct ExampleElementwiseOp<float, ck::bhalf_t>
// {
// __host__ __device__ void operator()(float& y, ck::bhalf_t& x) const
// {
// }
// };
struct AddReluAdd
{
......@@ -173,40 +160,109 @@ struct AddAdd
};
// C = A * B
// E = (C + D0) x D1
struct AddMultiply
{
template <typename E, typename C, typename D0, typename D1>
__host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
template <>
__host__ __device__ void operator()<half_t, half_t, half_t, half_t>(half_t& e,
const half_t& c,
const half_t& d0,
const half_t& d1) const
{
const half_t y = (c + d0) * d1;
e = y;
}
template <>
__host__ __device__ void operator()<half_t, float, half_t, half_t>(half_t& e,
const float& c,
const half_t& d0,
const half_t& d1) const
{
const half_t y = (type_convert<half_t>(c) + d0) * d1;
e = y;
}
template <>
__host__ __device__ void operator()<float, float, half_t, half_t>(float& e,
const float& c,
const half_t& d0,
const half_t& d1) const
{
const float y = (c + d0) * d1;
e = y;
}
};
// E = FastGelu(C + D0 + D1)
struct AddAddFastGelu
{
// Fast GeLU
// https://paperswithcode.com/method/gelu
// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
__host__ __device__ static constexpr float GetFastGeLU(float x)
template <typename E, typename C, typename D0, typename D1>
__host__ __device__ constexpr void
operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
template <>
__host__ __device__ constexpr void operator()<float, float, float, float>(float& e,
const float& c,
const float& d0,
const float& d1) const
{
const float u = 2.f * x * (0.035677f * x * x + 0.797885f);
const float emu = exp(-u);
const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
return x * cdf;
const float x = c + d0 + d1;
FastGelu{}.template operator()<float, float>(e, x);
}
template <typename T>
static inline constexpr bool is_valid_param_type_v =
std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|| std::is_same_v<T, ck::int4_t>
#endif
;
template <>
__host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
half_t& e, const half_t& c, const half_t& d0, const half_t& d1) const
{
const half_t x = c + d0 + d1;
template <typename E, typename C, typename D0, typename D1>
__host__ __device__ constexpr void
operator()(E& e, const C& c, const D0& d0, const D1& d1) const
ck::tensor_operation::element_wise::FastGelu{}.template operator()<half_t, half_t>(e, x);
}
template <>
__host__ __device__ constexpr void operator()<half_t, float, half_t, half_t>(
half_t& e, const float& c, const half_t& d0, const half_t& d1) const
{
const float x0_f = c + d0 + d1;
float x1_f = 0;
ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
x0_f);
e = type_convert<half_t>(x1_f);
}
template <>
__host__ __device__ constexpr void operator()<bhalf_t, float, bhalf_t, bhalf_t>(
bhalf_t& e, const float& c, const bhalf_t& d0, const bhalf_t& d1) const
{
static_assert(is_valid_param_type_v<E> && is_valid_param_type_v<C> &&
is_valid_param_type_v<D0> && is_valid_param_type_v<D1>);
const float x0_f = c + type_convert<float>(d0) + type_convert<float>(d1);
const float y =
GetFastGeLU(type_convert<float>(c) + type_convert<float>(d0) + type_convert<float>(d1));
float x1_f = 0;
e = type_convert<E>(y);
ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
x0_f);
e = type_convert<bhalf_t>(x1_f);
}
template <>
__host__ __device__ constexpr void operator()<int8_t, int32_t, int8_t, int8_t>(
int8_t& e, const int32_t& c, const int8_t& d0, const int8_t& d1) const
{
const float x0_f =
type_convert<float>(c) + type_convert<float>(d0) + type_convert<float>(d1);
float x1_f = 0;
ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
x0_f);
e = type_convert<int8_t>(x1_f);
}
};
......@@ -278,6 +334,40 @@ struct Normalize
double epsilon_;
};
// used by BatchNorm inference
// y = gamma * (x-mean) / sqrt(epsilon+variance) + beta
// The data type of mean and variance is used as AccDataType
struct NormalizeInInfer
{
NormalizeInInfer(double epsilon = 1e-4) : epsilon_(epsilon) {}
template <typename T1, typename T2, typename T3, typename T4>
__host__ __device__ constexpr void operator()(T1& y,
const T1& x,
const T2& mean,
const T2& variance,
const T3& gamma,
const T4& beta) const
{
static_assert(std::is_same<T2, float>::value || std::is_same<T2, double>::value,
"Data type is not supported by this operation!");
using ck::type_convert;
using ck::math::sqrt;
T2 tmp_x, tmp_y;
tmp_x = type_convert<T2>(x);
tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) *
type_convert<T2>(gamma) +
type_convert<T2>(beta);
y = type_convert<T1>(tmp_y);
};
double epsilon_;
};
template <typename Y, typename X>
struct UnaryTypeConvert;
......
include/ck/tensor_operation/gpu/element/quantization_operation.hpp
View file @ 8dd7156d
#pragma once
#include "ck/utility/data_type.hpp"
// #include "ck/utility/get_id.hpp"
namespace ck {
namespace tensor_operation {
namespace element_wise {
// Y = Sy * Qy
// W = Sw * Qw
// X = Sx * Qx
// B = Sb * Qb = Sw * Sx * Qb
// Where X, W, Y are float32, Qx, Qw, Qy are int8
// Sx, Sw, Sy are scale of x, w, y (float32), which is calculated from quantization range
// Qb is int32, scale of B is Sw * Sx for convenient
// Y = W @ X, where @ is convolution or matrix multiplication
// Sy * Qy = Sw * Qw @ Sx * Qx
// Qy = [(Sw*Sx)/Sy] * Qw @ Qx
// For Activation function which is piecewise linear function, such as relu, leaky relu ...etc
// Activation(Sy * Qy) = Sy * Activation(Qy)
template <typename Activation>
struct Activation_Mul_Clamp
{
// Convolution + Activation (piecewise linear function)
// If an activation is piecewise linear function, then Activation(Sy * Qy) = Sy * Activation(Qy)
// Z = Activation(Y) = Activation(W @ X)
// Sz * Qz = Activation(Sy * Qy)
// Qz = Sy / Sz * Activation(Qy) = (Sw * Sx / Sz) * Activation(Qw @ Qx)
// requantScale_ = Sw * Sx / Sz
Activation_Mul_Clamp(float requantScale, Activation activationOp)
: requantScale_(requantScale), activationOp_(activationOp)
{
......@@ -17,26 +38,66 @@ struct Activation_Mul_Clamp
__host__ __device__ constexpr void operator()(int8_t& y, const int32_t& x) const
{
float x_fp32 = ck::type_convert<float>(x);
activationOp_(x_fp32, x_fp32);
float y_fp32 = math::clamp(requantScale_ * x_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
float y_fp32 = ck::type_convert<float>(x);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
__device__ constexpr void operator()(int32_t& y, const int32_t& x) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
__host__ __device__ constexpr void operator()(float& y, const int32_t& x) const
__host__ constexpr void operator()(float& y, const float& x) const
{
// We might type_convert to int8 after lambda in someplace
float x_fp32 = ck::type_convert<float>(x);
activationOp_(x_fp32, x_fp32);
y = math::clamp(requantScale_ * x_fp32, -128.f, 127.f);
// CAUSION - We might float in & float out in reference code
activationOp_(y, x);
y = math::clamp(requantScale_ * y, -128.f, 127.f);
}
float requantScale_;
Activation activationOp_;
};
// For Activation function which is non piecewise linear function, such as TanH, Sigmoid ...etc
// If an activation is not piecewise linear function
// then Activation(Sy * Qy) != Sy * Activation(Qy)
template <typename Activation>
struct Mul_Activation_Mul_Clamp
{
// Convolution + Activation (non piecewise linear function)
// Z = Activation(Y) = Activation(W @ X)
// Sz * Qz = Activation(Sy * Qy)
// Qz = S1 * Activation[Sacc * (Qw @ Qx)]
// Where S1 = 1 / Sz, Sacc = Sw * Sx
Mul_Activation_Mul_Clamp(float scale_z_inv, float scaleAcc, Activation activationOp)
: scale_z_inv_(scale_z_inv), scaleAcc_(scaleAcc), activationOp_(activationOp)
{
}
__host__ __device__ constexpr void operator()(int8_t& y, const int32_t& x) const
{
float y_fp32 = ck::type_convert<float>(x);
y_fp32 = scaleAcc_ * y_fp32;
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
float scale_z_inv_;
float scaleAcc_;
Activation activationOp_;
};
// Conv Perchannel quantization + Activation function which is piecewise linear function, such as
// relu, leaky relu ...etc
// Activation(Sy * Qy) = Sy * Activation(Qy)
template <typename Activation>
struct Activation_Mul2_Clamp
{
......@@ -51,13 +112,35 @@ struct Activation_Mul2_Clamp
y = ck::type_convert<int8_t>(y_fp32);
}
__device__ constexpr void
operator()(int32_t& y, const int32_t& x, const float& requantScale) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
Activation activationOp_;
};
// For Activation function which is piecewise linear function, such as relu, leaky relu ...etc
// Activation(Sy * Qy) = Sy * Activation(Qy)
template <typename Activation>
struct Add_Activation_Mul_Clamp
{
// Convolution + bias
// Let Bias = B = Sw * Sx * Qb
// Where Qb is int32
// Y = W @ X + B
// Sy * Qy = Sw * Qw @ Sx * Qx + Sw * Sx * Qb
// Qy = [(Sw*Sx)/Sy] * (Qw @ Qx + Qb)
// For activation, Z = Activaiton(Y)
// Sz * Qz = Activation(Sy * Qy)
// Qz = Sy / Sz * Activation(Qy) = [(Sw*Sx)/Sz] * Activation(Qw @ Qx + Qb)
Add_Activation_Mul_Clamp(float requantScale, Activation activationOp)
: requantScale_(requantScale), activationOp_(activationOp)
{
......@@ -72,6 +155,17 @@ struct Add_Activation_Mul_Clamp
y = ck::type_convert<int8_t>(y_fp32);
}
__host__ __device__ constexpr void
operator()(int32_t& y, const int32_t& x, const int32_t& bias) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x + bias);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
float requantScale_;
Activation activationOp_;
};
......@@ -92,15 +186,33 @@ struct Add_Activation_Mul2_Clamp
y = ck::type_convert<int8_t>(y_fp32);
}
__host__ __device__ constexpr void
operator()(int32_t& y, const int32_t& x, const int32_t& bias, const float& requantScale) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x + bias);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
Activation activationOp_;
};
// For Activation function which is non piecewise linear function, such as TanH, Sigmoid ...etc
// If an activation is not piecewise linear function
// then Activation(Sy * Qy) != Sy * Activation(Qy)
template <typename Activation>
struct Add_Mul_Activation_Mul_Clamp
{
Add_Mul_Activation_Mul_Clamp(float requantScale1, float requantScale2, Activation activationOp)
: requantScale1_(requantScale1), requantScale2_(requantScale2), activationOp_(activationOp)
// Convolution + Activation (non piecewise linear function)
// Z = Activation(Y) = Activation(W @ X + B)
// Sz * Qz = Activation(Sy * Qy)
// Qz = S1 * Activation[Sacc * (Qw @ Qx + Qb)]
// Where S1 = 1 / Sz, Sacc = Sw * Sx
Add_Mul_Activation_Mul_Clamp(float scale_z_inv, float scaleAcc, Activation activationOp)
: scale_z_inv_(scale_z_inv), scaleAcc_(scaleAcc), activationOp_(activationOp)
{
}
......@@ -108,14 +220,64 @@ struct Add_Mul_Activation_Mul_Clamp
operator()(int8_t& y, const int32_t& x, const int32_t& bias) const
{
float y_fp32 = ck::type_convert<float>(x + bias);
y_fp32 = requantScale1_ * y_fp32;
y_fp32 = scaleAcc_ * y_fp32;
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
__host__ __device__ constexpr void
operator()(int32_t& y, const int32_t& x, const int32_t& bias) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x + bias);
y_fp32 = scaleAcc_ * y_fp32;
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
float scale_z_inv_;
float scaleAcc_;
Activation activationOp_;
};
// Conv Perchannel quantization + Activation function which is non piecewise linear function,
// such as TanH, Sigmoid ...etc
// If an activation is not piecewise linear function
// then Activation(Sy *Qy) != Sy * Activation(Qy)
template <typename Activation>
struct Add_Mul2_Activation_Mul_Clamp
{
Add_Mul2_Activation_Mul_Clamp(float scale_z_inv, Activation activationOp)
: scale_z_inv_(scale_z_inv), activationOp_(activationOp)
{
}
__host__ __device__ constexpr void
operator()(int8_t& y, const int32_t& x, const int32_t& bias, const float& scaleAcc) const
{
float y_fp32 = ck::type_convert<float>(x + bias);
y_fp32 = scaleAcc * y_fp32;
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale2_ * y_fp32, -128.f, 127.f);
y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
float requantScale1_;
float requantScale2_;
__host__ __device__ constexpr void
operator()(int32_t& y, const int32_t& x, const int32_t& bias, const float& scaleAcc) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x + bias);
y_fp32 = scaleAcc * y_fp32;
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
float scale_z_inv_;
Activation activationOp_;
};
......
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/math.hpp"
#include "ck/utility/math_v2.hpp"
#include "ck/utility/type_convert.hpp"
namespace ck {
namespace tensor_operation {
namespace element_wise {
#if CK_WORKAROUND_SWDEV_383542
extern "C" __device__ float __ocml_native_recip_f32(float);
#endif
struct PassThrough
{
template <typename Y, typename X>
......@@ -52,6 +57,12 @@ struct PassThrough
y = type_convert<bhalf_t>(x);
}
template <>
__host__ __device__ void operator()<bhalf_t, half_t>(bhalf_t& y, const half_t& x) const
{
y = type_convert<bhalf_t>(x);
}
template <>
__host__ __device__ void operator()<int8_t, int8_t>(int8_t& y, const int8_t& x) const
{
......@@ -71,6 +82,36 @@ struct PassThrough
y = x;
}
#endif
template <>
__host__ __device__ void operator()<f8_t, f8_t>(f8_t& y, const f8_t& x) const
{
y = x;
}
template <>
__host__ __device__ void operator()<float, f8_t>(float& y, const f8_t& x) const
{
y = type_convert<float>(x);
}
template <>
__host__ __device__ void operator()<f8_t, float>(f8_t& y, const float& x) const
{
y = type_convert<f8_t>(x);
}
template <>
__host__ __device__ void operator()<half_t, f8_t>(half_t& y, const f8_t& x) const
{
y = type_convert<half_t>(x);
}
template <>
__host__ __device__ void operator()<f8_t, half_t>(f8_t& y, const half_t& x) const
{
y = type_convert<f8_t>(x);
}
};
struct UnaryConvert
......@@ -82,6 +123,40 @@ struct UnaryConvert
}
};
struct ConvertBF16RTN
{
// convert to bf16 using round to nearest (rtn)
template <typename Y, typename X>
__host__ __device__ void operator()(Y& y, const X& x) const
{
// check Y datatype
static_assert(is_same<Y, bhalf_t>::value, "Data type is not supported by this operation!");
// check X datatype
static_assert(is_same<X, float>::value || is_same<X, half_t>::value,
"Data type is not supported by this operation!");
y = bf16_convert_rtn<Y>(x);
}
};
struct ConvertF8SR
{
// convert to fp8 using stochastic rounding (SR)
template <typename Y, typename X>
__host__ __device__ void operator()(Y& y, const X& x) const
{
// check Y datatype
static_assert(is_same<Y, f8_t>::value, "Data type is not supported by this operation!");
// check X datatype
static_assert(is_same<X, float>::value || is_same<X, half_t>::value,
"Data type is not supported by this operation!");
y = f8_convert_sr<Y>(x);
}
};
struct Scale
{
__host__ __device__ Scale(float scale) : scale_(scale) {}
......@@ -95,6 +170,12 @@ struct Scale
y = scale_ * x;
};
template <>
__host__ __device__ void operator()<double, double>(double& y, const double& x) const
{
y = scale_ * x;
};
__host__ __device__ auto Value() const { return scale_; }
float scale_;
......@@ -196,36 +277,83 @@ struct Relu
}
};
// Y = FastGelu(X)
// Fast GeLU
// https://paperswithcode.com/method/gelu
// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
// host code use higher accuracy "exp" and "div"
// gpu code use lower accuracy "__expf" and "rcp" function
struct FastGelu
{
// Fast GeLU
// https://paperswithcode.com/method/gelu
// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
__host__ __device__ static constexpr float GetFastGeLU(float x)
template <typename Y, typename X>
__host__ void operator()(Y& y, const X& x) const;
template <typename Y, typename X>
__device__ void operator()(Y& y, const X& x) const;
template <>
__host__ void operator()<float, float>(float& y, const float& x) const
{
const float u = 2.f * x * (0.035677f * x * x + 0.797885f);
const float emu = exp(-u);
const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
return x * cdf;
y = x * cdf;
}
template <typename T>
static inline constexpr bool is_valid_param_type_v =
std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|| std::is_same_v<T, ck::int4_t>
// device code, use lower precision "__expf" and "rcp"
template <>
__device__ void operator()<float, float>(float& y, const float& x) const
{
const float u = 2.f * x * (0.035677f * x * x + 0.797885f);
const float emu = __expf(-u);
#if !CK_WORKAROUND_SWDEV_383542
const float cdf = 0.5f + 0.5f * (2.f * __frcp_rn(1.f + emu) - 1.f);
#else
const float cdf = 0.5f + 0.5f * (2.f * __ocml_native_recip_f32(1.f + emu) - 1.f);
#endif
;
template <typename Y, typename X>
__host__ __device__ void operator()(Y& y, const X& x) const
y = x * cdf;
}
template <>
__host__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
{
static_assert(is_valid_param_type_v<Y> && is_valid_param_type_v<X>);
float y_f;
this->operator()<float, float>(y_f, type_convert<float>(x));
const float tmp_y = GetFastGeLU(type_convert<float>(x));
y = type_convert<Y>(tmp_y);
y = type_convert<half_t>(y_f);
}
template <>
__device__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
{
float y_f;
this->operator()<float, float>(y_f, type_convert<float>(x));
y = type_convert<half_t>(y_f);
}
template <>
__host__ void operator()<half_t, float>(half_t& y, const float& x) const
{
float y_f;
this->operator()<float, float>(y_f, x);
y = type_convert<half_t>(y_f);
}
template <>
__device__ void operator()<half_t, float>(half_t& y, const float& x) const
{
float y_f;
this->operator()<float, float>(y_f, x);
y = type_convert<half_t>(y_f);
}
};
......@@ -261,8 +389,36 @@ struct Sigmoid
y = 1 / (ck::type_convert<T>(1) + exp(-x));
};
};
int32_t divider_ = 1;
struct TanH
{
template <typename T>
__host__ __device__ void operator()(T& y, const T& x) const
{
static_assert(is_same<T, float>::value || is_same<T, double>::value ||
is_same<T, ck::half_t>::value,
"Data type is not supported by this operation!");
y = ck::math::tanh(x);
};
};
struct Swish
{
Swish(float beta = 1.0f) : beta_(beta) {}
template <typename T>
__host__ __device__ void operator()(T& y, const T& x) const
{
static_assert(is_same<T, float>::value || is_same<T, double>::value ||
is_same<T, ck::half_t>::value,
"Data type is not supported by this operation!");
y = x / (ck::type_convert<T>(1) + ck::math::exp(-beta_ * x));
};
float beta_ = 1.0f;
};
} // namespace element_wise
......
include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp 0 → 100644
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/math_v2.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/utility/workgroup_synchronization.hpp"
namespace ck {
template <typename GridwiseMultiblockBatchNormForward_,
typename XDataType,
typename YDataType,
typename AccDataType,
typename ScaleDataType,
typename BiasDataType,
typename MeanVarDataType,
typename YElementwiseOp,
typename XYGridDesc_M_K,
typename MeanVarCountGridDesc_M_G,
typename MeanVarCountGridDesc_M_K,
typename ScaleBiasGridDesc_M,
typename MeanVarGridDesc_M,
typename GetReduceCountPerThreadFunctor>
__global__ void kernel_multiblock_batchnorm_forward(
const XYGridDesc_M_K x_grid_desc_m_k,
const XYGridDesc_M_K y_grid_desc_m_k,
const MeanVarCountGridDesc_M_G mean_var_count_grid_desc_m_g,
const MeanVarCountGridDesc_M_K mean_var_count_grid_desc_m_k,
const ScaleBiasGridDesc_M scale_grid_desc_m,
const ScaleBiasGridDesc_M bias_grid_desc_m,
const MeanVarGridDesc_M mean_var_grid_desc_m,
const GetReduceCountPerThreadFunctor get_reduce_count_per_thread,
index_t num_k_block_tile_iteration,
AccDataType epsilon,
const XDataType* const __restrict__ p_x,
MeanVarDataType* const __restrict__ p_welford_mean,
MeanVarDataType* const __restrict__ p_welford_variance,
int32_t* const __restrict__ p_welford_count,
int32_t* const __restrict__ p_control,
const ScaleDataType* const __restrict__ p_scale,
const BiasDataType* const __restrict__ p_bias,
const YElementwiseOp y_elementwise_op,
YDataType* const __restrict__ p_y,
bool updateMovingAverage,
AccDataType averageFactor,
MeanVarDataType* const __restrict__ resultRunningMean,
MeanVarDataType* const __restrict__ resultRunningVariance,
bool saveMeanInvVariance,
MeanVarDataType* const __restrict__ resultSaveMean,
MeanVarDataType* const __restrict__ resultSaveInvVariance)
{
GridwiseMultiblockBatchNormForward_::Run(x_grid_desc_m_k,
y_grid_desc_m_k,
mean_var_count_grid_desc_m_g,
mean_var_count_grid_desc_m_k,
scale_grid_desc_m,
bias_grid_desc_m,
mean_var_grid_desc_m,
get_reduce_count_per_thread,
num_k_block_tile_iteration,
epsilon,
p_x,
p_welford_mean,
p_welford_variance,
p_welford_count,
p_control,
p_scale,
p_bias,
y_elementwise_op,
p_y,
updateMovingAverage,
averageFactor,
resultRunningMean,
resultRunningVariance,
saveMeanInvVariance,
resultSaveMean,
resultSaveInvVariance);
};
template <typename XDataType,
typename YDataType,
typename AccDataType,
typename ScaleDataType,
typename BiasDataType,
typename MeanVarDataType,
typename YElementwiseOp,
typename XYGridDesc_M_K,
typename MeanVarCountGridDesc_M_G,
typename MeanVarCountGridDesc_M_K,
typename ScaleBiasGridDesc_M,
typename MeanVarGridDesc_M,
typename GetReduceCountPerThreadFunctor,
index_t BlockSize,
index_t MThreadClusterSize,
index_t KThreadClusterSize,
index_t MThreadSliceSize,
index_t KThreadSliceSize,
index_t XSrcYDstVectorDim,
index_t XSrcVectorSize,
index_t YDstVectorSize,
index_t ScaleSrcVectorSize,
index_t BiasSrcVectorSize,
index_t MeanVarSrcDstVectorSize>
struct GridwiseMultiblockBatchNormForward
{
static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
(XSrcYDstVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
"Invalid thread slice sizes and/or vector sizes configuration, please check!");
static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
(XSrcYDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
"Invalid thread slice sizes and/or vector sizes configuration, please check!");
static constexpr bool reorder_thread_cluster = (XSrcYDstVectorDim == 0);
using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
using ThreadBufferDimAccessOrder =
typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
using ThreadClusterArrangeOrder =
typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
static constexpr auto thread_cluster_desc =
make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
using ThreadReduceDstDesc_M =
decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
using ThreadReduceSrcDesc_M_1 = decltype(
make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
using ThreadwiseWelford1 =
ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
using ThreadwiseWelford2 =
ThreadwiseWelfordMerge<AccDataType, ThreadReduceSrcDesc_M_1, ThreadReduceDstDesc_M>;
using BlockwiseWelford1 = BlockwiseWelford<AccDataType,
BlockSize,
ThreadClusterLengths_M_K,
ThreadClusterArrangeOrder,
false>;
using BlockwiseWelford2 = BlockwiseWelford<AccDataType,
BlockSize,
ThreadClusterLengths_M_K,
ThreadClusterArrangeOrder,
true>;
using PassThroughOp = tensor_operation::element_wise::PassThrough;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
__device__ static void Run(const XYGridDesc_M_K& x_grid_desc_m_k,
const XYGridDesc_M_K& y_grid_desc_m_k,
const MeanVarCountGridDesc_M_G& mean_var_count_grid_desc_m_g,
const MeanVarCountGridDesc_M_K& mean_var_count_grid_desc_m_k,
const ScaleBiasGridDesc_M& scale_grid_desc_m,
const ScaleBiasGridDesc_M& bias_grid_desc_m,
const MeanVarGridDesc_M& mean_var_grid_desc_m,
const GetReduceCountPerThreadFunctor& get_reduce_count_per_thread,
index_t num_k_block_tile_iteration,
AccDataType epsilon,
const XDataType* const __restrict__ p_x,
MeanVarDataType* const __restrict__ p_welford_mean,
MeanVarDataType* const __restrict__ p_welford_variance,
int32_t* const __restrict__ p_welford_count,
int32_t* const __restrict__ p_control,
const ScaleDataType* const __restrict__ p_scale,
const BiasDataType* const __restrict__ p_bias,
const YElementwiseOp y_elementwise_op,
YDataType* const __restrict__ p_y,
bool updateMovingAverage,
AccDataType averageFactor,
MeanVarDataType* const __restrict__ resultRunningMean,
MeanVarDataType* const __restrict__ resultRunningVariance,
bool saveMeanInvVariance,
MeanVarDataType* const __restrict__ resultSaveMean,
MeanVarDataType* const __restrict__ resultSaveInvVariance)
{
using ck::math::sqrt;
const index_t blkgroup_size = mean_var_count_grid_desc_m_g.GetLength(I1);
const index_t thread_local_id = get_thread_local_1d_id();
const index_t block_global_id = get_block_1d_id();
const index_t blkgroup_id = block_global_id / blkgroup_size;
const index_t block_local_id = block_global_id % blkgroup_size;
if(block_local_id == 0)
gms_init(BlockSize / warpSize * blkgroup_size, &p_control[blkgroup_id * 2]);
const auto thread_cluster_idx =
thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
const auto thread_m_cluster_id = thread_cluster_idx[I0];
const auto thread_k_cluster_id = thread_cluster_idx[I1];
using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, KThreadSliceSize>;
using ThreadBufferLengths_M = Sequence<MThreadSliceSize>;
using ThreadBufferLengths_M_1 = Sequence<MThreadSliceSize, 1>;
constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
constexpr auto thread_buffer_desc_m =
make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
constexpr auto thread_buffer_desc_m_1 = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
x_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true> count_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
tmp_mean_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
tmp_var_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true> tmp_count_thread_buf;
const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
AccDataType,
XYGridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
ThreadBufferDimAccessOrder,
XSrcYDstVectorDim,
XSrcVectorSize,
1,
true>(
x_grid_desc_m_k,
make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
block_local_id * reduceSizePerBlock +
thread_k_cluster_id * KThreadSliceSize));
constexpr auto xy_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileSize);
const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_x, x_grid_desc_m_k.GetElementSpaceSize());
// Step 1: each workgroup does local welford reduction
auto threadwise_welford_1 = ThreadwiseWelford1();
threadwise_welford_1.max_count_ =
get_reduce_count_per_thread(block_local_id, thread_k_cluster_id);
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
var_thread_buf(I) = type_convert<AccDataType>(0.0f);
});
for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
{
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf);
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, xy_copy_fwd_step_m_k);
threadwise_welford_1.Run(x_thread_buf, mean_thread_buf, var_thread_buf);
}
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
if constexpr(I > 0)
block_sync_lds();
count_thread_buf(I) = threadwise_welford_1.cur_count_;
BlockwiseWelford1::Run(mean_thread_buf(I), var_thread_buf(I), count_thread_buf(I));
});
// Step 2: each workgroup writes its local welford result to workspace memory
auto mean_global_val_buf =
make_dynamic_buffer<AddressSpaceEnum::Global, AmdBufferCoherenceEnum::GLC>(
p_welford_mean, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
auto var_global_val_buf =
make_dynamic_buffer<AddressSpaceEnum::Global, AmdBufferCoherenceEnum::GLC>(
p_welford_variance, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
auto count_global_val_buf =
make_dynamic_buffer<AddressSpaceEnum::Global, AmdBufferCoherenceEnum::GLC>(
p_welford_count, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
auto threadwise_mean_var_store_m_g =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
MeanVarDataType,
decltype(thread_buffer_desc_m_1),
MeanVarCountGridDesc_M_G,
PassThroughOp,
ThreadBufferLengths_M_1,
Sequence<0, 1>,
0,
1,
InMemoryDataOperationEnum::Set,
1,
true>(
mean_var_count_grid_desc_m_g,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
block_local_id),
PassThroughOp{});
auto threadwise_count_store_m_g =
ThreadwiseTensorSliceTransfer_v1r3<int32_t,
int32_t,
decltype(thread_buffer_desc_m_1),
MeanVarCountGridDesc_M_G,
PassThroughOp,
ThreadBufferLengths_M_1,
Sequence<0, 1>,
0,
1,
InMemoryDataOperationEnum::Set,
1,
true>(
mean_var_count_grid_desc_m_g,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
block_local_id),
PassThroughOp{});
if(thread_k_cluster_id == 0)
{
threadwise_mean_var_store_m_g.Run(thread_buffer_desc_m_1,
make_tuple(I0, I0),
mean_thread_buf,
mean_var_count_grid_desc_m_g,
mean_global_val_buf);
threadwise_mean_var_store_m_g.Run(thread_buffer_desc_m_1,
make_tuple(I0, I0),
var_thread_buf,
mean_var_count_grid_desc_m_g,
var_global_val_buf);
threadwise_count_store_m_g.Run(thread_buffer_desc_m_1,
make_tuple(I0, I0),
count_thread_buf,
mean_var_count_grid_desc_m_g,
count_global_val_buf);
};
gms_barrier(&p_control[blkgroup_id * 2]);
if(block_local_id == 0)
gms_reset(&p_control[blkgroup_id * 2]);
// Step 3: each workgroup reads welford results from workspace memory and does final welford
// reduction
auto threadwise_mean_var_load_m_k =
ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
AccDataType,
MeanVarCountGridDesc_M_K,
decltype(thread_buffer_desc_m_1),
ThreadBufferLengths_M_1,
Sequence<0, 1>,
0,
1,
1,
true>(
mean_var_count_grid_desc_m_k,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * 1));
auto threadwise_count_load_m_k =
ThreadwiseTensorSliceTransfer_v2<int32_t,
int32_t,
MeanVarCountGridDesc_M_K,
decltype(thread_buffer_desc_m_1),
ThreadBufferLengths_M_1,
Sequence<0, 1>,
0,
1,
1,
true>(
mean_var_count_grid_desc_m_k,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * 1));
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
var_thread_buf(I) = type_convert<AccDataType>(0.0f);
count_thread_buf(I) = 0;
});
constexpr auto mean_var_count_read_fwd_step_m_k = make_multi_index(0, KThreadClusterSize);
int32_t reducedSize = 0;
while(reducedSize < blkgroup_size)
{
threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
mean_global_val_buf,
thread_buffer_desc_m_1,
make_tuple(I0, I0),
tmp_mean_thread_buf);
threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
var_global_val_buf,
thread_buffer_desc_m_1,
make_tuple(I0, I0),
tmp_var_thread_buf);
threadwise_count_load_m_k.Run(mean_var_count_grid_desc_m_k,
count_global_val_buf,
thread_buffer_desc_m_1,
make_tuple(I0, I0),
tmp_count_thread_buf);
ThreadwiseWelford2::Run(tmp_mean_thread_buf,
tmp_var_thread_buf,
tmp_count_thread_buf,
mean_thread_buf,
var_thread_buf,
count_thread_buf);
reducedSize += KThreadClusterSize;
threadwise_mean_var_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
mean_var_count_read_fwd_step_m_k);
threadwise_count_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
mean_var_count_read_fwd_step_m_k);
};
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
if constexpr(I > 0)
block_sync_lds();
BlockwiseWelford2::Run(mean_thread_buf(I), var_thread_buf(I), count_thread_buf(I));
});
// Step 4: do normalization using the mean/variance
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> scale_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> bias_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
y_thread_buf;
auto threadwise_y_store =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
YDataType,
decltype(thread_buffer_desc_m_k),
XYGridDesc_M_K,
YElementwiseOp,
ThreadBufferLengths_M_K,
ThreadBufferDimAccessOrder,
XSrcYDstVectorDim,
YDstVectorSize,
InMemoryDataOperationEnum::Set,
1,
true>(
y_grid_desc_m_k,
make_multi_index(
blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
block_local_id * reduceSizePerBlock + thread_k_cluster_id * KThreadSliceSize),
y_elementwise_op);
auto threadwise_scale_load =
ThreadwiseTensorSliceTransfer_v2<ScaleDataType,
AccDataType,
ScaleBiasGridDesc_M,
decltype(thread_buffer_desc_m),
ThreadBufferLengths_M,
Sequence<0>,
0,
ScaleSrcVectorSize,
1,
true>(
scale_grid_desc_m,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize));
auto threadwise_bias_load = ThreadwiseTensorSliceTransfer_v2<BiasDataType,
AccDataType,
ScaleBiasGridDesc_M,
decltype(thread_buffer_desc_m),
ThreadBufferLengths_M,
Sequence<0>,
0,
BiasSrcVectorSize,
1,
true>(
bias_grid_desc_m,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize));
const auto scale_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_scale, scale_grid_desc_m.GetElementSpaceSize());
const auto bias_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_bias, bias_grid_desc_m.GetElementSpaceSize());
auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_y, y_grid_desc_m_k.GetElementSpaceSize());
threadwise_scale_load.Run(scale_grid_desc_m,
scale_global_val_buf,
thread_buffer_desc_m,
make_tuple(I0),
scale_thread_buf);
threadwise_bias_load.Run(bias_grid_desc_m,
bias_global_val_buf,
thread_buffer_desc_m,
make_tuple(I0),
bias_thread_buf);
threadwise_x_load.SetSrcSliceOrigin(
x_grid_desc_m_k,
make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
block_local_id * reduceSizePerBlock +
thread_k_cluster_id * KThreadSliceSize));
for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
{
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf);
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
AccDataType multiplier =
scale_thread_buf[Number<iM>{}] / sqrt(var_thread_buf[iM] + epsilon);
AccDataType fused_mean_bias =
bias_thread_buf[Number<iM>{}] - mean_thread_buf[iM] * multiplier;
static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
constexpr auto offset =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
// normalize
y_thread_buf(Number<offset>{}) =
x_thread_buf[Number<offset>{}] * multiplier + fused_mean_bias;
});
});
threadwise_y_store.Run(thread_buffer_desc_m_k,
make_tuple(I0, I0),
y_thread_buf,
y_grid_desc_m_k,
y_global_val_buf);
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, xy_copy_fwd_step_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, xy_copy_fwd_step_m_k);
}
// Step 5: update the moving average of mean and variance (optional)
if(updateMovingAverage && block_local_id == 0 && thread_k_cluster_id == 0)
{
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
running_mean_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
running_var_thread_buf;
auto running_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
resultRunningMean, mean_var_grid_desc_m.GetElementSpaceSize());
auto running_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
resultRunningVariance, mean_var_grid_desc_m.GetElementSpaceSize());
auto threadwise_mean_var_load =
ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
AccDataType,
MeanVarGridDesc_M,
decltype(thread_buffer_desc_m),
ThreadBufferLengths_M,
Sequence<0>,
0,
MeanVarSrcDstVectorSize,
1,
true>(
mean_var_grid_desc_m,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize));
threadwise_mean_var_load.Run(mean_var_grid_desc_m,
running_mean_global_buf,
thread_buffer_desc_m,
make_tuple(I0),
running_mean_thread_buf);
threadwise_mean_var_load.Run(mean_var_grid_desc_m,
running_var_global_buf,
thread_buffer_desc_m,
make_tuple(I0),
running_var_thread_buf);
AccDataType oneMinusAverageFactor = type_convert<AccDataType>(1.0) - averageFactor;
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
running_mean_thread_buf(I) = running_mean_thread_buf[I] * oneMinusAverageFactor +
mean_thread_buf[I] * averageFactor;
running_var_thread_buf(I) = running_var_thread_buf[I] * oneMinusAverageFactor +
var_thread_buf[I] * averageFactor;
});
auto threadwise_mean_var_store =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
MeanVarDataType,
decltype(thread_buffer_desc_m),
MeanVarGridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>,
0,
MeanVarSrcDstVectorSize,
InMemoryDataOperationEnum::Set,
1,
true>(
mean_var_grid_desc_m,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
threadwise_mean_var_store.Run(thread_buffer_desc_m,
make_tuple(I0),
running_mean_thread_buf,
mean_var_grid_desc_m,
running_mean_global_buf);
threadwise_mean_var_store.Run(thread_buffer_desc_m,
make_tuple(I0),
running_var_thread_buf,
mean_var_grid_desc_m,
running_var_global_buf);
};
// Step 6: save mean and inv-variance (optional)
if(saveMeanInvVariance && block_local_id == 0 && thread_k_cluster_id == 0)
{
auto result_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
resultSaveMean, mean_var_grid_desc_m.GetElementSpaceSize());
auto result_inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
resultSaveInvVariance, mean_var_grid_desc_m.GetElementSpaceSize());
// calculate inv-variance as 1/sqrt(epsilon+variance), stored in place of variance
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
var_thread_buf(I) =
type_convert<AccDataType>(1.0f) / sqrt(epsilon + var_thread_buf[I]);
});
auto threadwise_mean_inv_var_store =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
MeanVarDataType,
decltype(thread_buffer_desc_m),
MeanVarGridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>,
0,
MeanVarSrcDstVectorSize,
InMemoryDataOperationEnum::Set,
1,
true>(
mean_var_grid_desc_m,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
make_tuple(I0),
mean_thread_buf,
mean_var_grid_desc_m,
result_mean_global_buf);
threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
make_tuple(I0),
var_thread_buf,
mean_var_grid_desc_m,
result_inv_var_global_buf);
};
}
}; // namespace ck
} // namespace ck
include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......
include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -161,7 +161,7 @@ struct GridwiseMultiblockWelfordFirstHalf
PassThroughOp,
ThreadBufferLengths_M_1,
Sequence<0, 1>,
1,
0,
1,
InMemoryDataOperationEnum::Set,
1,
......@@ -180,7 +180,7 @@ struct GridwiseMultiblockWelfordFirstHalf
PassThroughOp,
ThreadBufferLengths_M_1,
Sequence<0, 1>,
1,
0,
1,
InMemoryDataOperationEnum::Set,
1,
......
include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final.hpp → include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -33,7 +33,6 @@ __global__ void kernel_welford_second_half_batchnorm_forward_final(
const MeanVarGridDesc_M mean_var_grid_desc_m,
index_t blkgroup_size,
index_t num_xy_k_block_tile_iteration,
index_t num_mean_var_count_k_block_tile_iteration,
AccDataType epsilon,
const MeanVarDataType* const __restrict__ p_in_welford_mean,
const MeanVarDataType* const __restrict__ p_in_welford_variance,
......@@ -59,7 +58,6 @@ __global__ void kernel_welford_second_half_batchnorm_forward_final(
mean_var_grid_desc_m,
blkgroup_size,
num_xy_k_block_tile_iteration,
num_mean_var_count_k_block_tile_iteration,
epsilon,
p_in_welford_mean,
p_in_welford_variance,
......@@ -152,7 +150,6 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
const MeanVarGridDesc_M& mean_var_grid_desc_m,
index_t blkgroup_size,
index_t num_xy_k_block_tile_iteration,
index_t num_mean_var_count_k_block_tile_iteration,
AccDataType epsilon,
const MeanVarDataType* const __restrict__ p_in_welford_mean,
const MeanVarDataType* const __restrict__ p_in_welford_variance,
......@@ -223,7 +220,7 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
decltype(thread_buffer_desc_m_1),
ThreadBufferLengths_M_1,
Sequence<0, 1>,
1,
0,
1,
1,
true>(
......@@ -239,7 +236,7 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
decltype(thread_buffer_desc_m_1),
ThreadBufferLengths_M_1,
Sequence<0, 1>,
1,
0,
1,
1,
true>(
......@@ -257,9 +254,6 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
const auto welford_count_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_in_welford_count, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
constexpr auto mean_var_count_thread_copy_step_m_k =
make_multi_index(0, KThreadClusterSize * 1);
// Step 1: do final welford reduction to get mean and variance
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
......@@ -268,8 +262,11 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
welford_count_thread_buf(I) = 0;
});
for(index_t reducedTiles = 0; reducedTiles < num_mean_var_count_k_block_tile_iteration;
++reducedTiles)
constexpr auto mean_var_count_thread_copy_step_m_k =
make_multi_index(0, KThreadClusterSize);
int32_t reducedSize = 0;
while(reducedSize < blkgroup_size)
{
threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
welford_mean_global_val_buf,
......@@ -296,6 +293,8 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
welford_var_thread_buf,
welford_count_thread_buf);
reducedSize += KThreadClusterSize;
threadwise_mean_var_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
mean_var_count_thread_copy_step_m_k);
threadwise_count_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
......
include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View file @ 8dd7156d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -109,30 +109,57 @@ struct BlockToCTileMap_M00_N0_M01
// Rows of column-vectors
// This C-tile map dynamically adjusts M01 when C-tile index is out of range
template <index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N>
struct BlockToCTileMap_M00_N0_M01Adapt
template <index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N = void>
struct BlockToCTileMap_M00_N0_M01Adapt;
template <index_t MPerBlock, index_t NPerBlock>
struct BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
__host__ __device__ BlockToCTileMap_M00_N0_M01Adapt() = default;
__host__ __device__ BlockToCTileMap_M00_N0_M01Adapt(const BlockToCTileMap_M00_N0_M01Adapt&) =
default;
__host__ __device__ BlockToCTileMap_M00_N0_M01Adapt(BlockToCTileMap_M00_N0_M01Adapt&&) =
default;
__host__ __device__ BlockToCTileMap_M00_N0_M01Adapt&
operator=(const BlockToCTileMap_M00_N0_M01Adapt&) = default;
__host__ __device__ BlockToCTileMap_M00_N0_M01Adapt&
operator=(BlockToCTileMap_M00_N0_M01Adapt&&) = default;
__host__ __device__ BlockToCTileMap_M00_N0_M01Adapt(index_t M, index_t N, index_t M01 = 8)
: M_(M), N_(N), M01_(M01)
{
}
template <typename CGridDesc_M_N>
__host__ __device__ BlockToCTileMap_M00_N0_M01Adapt(const CGridDesc_M_N& c_grid_desc_m_n,
index_t M01 = 8)
: M01_(M01), c_grid_desc_m_n_(c_grid_desc_m_n)
: BlockToCTileMap_M00_N0_M01Adapt(
c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1), M01)
{
}
__host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
__host__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
{
const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
const auto M0 = math::integer_divide_ceil(M, MPerBlock);
const auto N0 = math::integer_divide_ceil(N, NPerBlock);
const index_t grid_size = M0 * N0;
return M0 * N0;
}
return grid_size;
template <typename CGridDesc_M_N>
__host__ static constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n)
{
return CalculateGridSize(c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1));
}
template <typename CGridDesc_M_N>
__host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
{
return true;
}
template <typename TopIdx>
......@@ -140,8 +167,8 @@ struct BlockToCTileMap_M00_N0_M01Adapt
{
auto block_1d_id = idx_top[I0];
const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I0), MPerBlock);
const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I1), NPerBlock);
const auto M0 = math::integer_divide_ceil(M_, MPerBlock);
const auto N0 = math::integer_divide_ceil(N_, NPerBlock);
block_1d_id = block_1d_id % (M0 * N0); // swallow batch index
......@@ -154,6 +181,50 @@ struct BlockToCTileMap_M00_N0_M01Adapt
index_t idx_M01 = idx_M0 % M01_;
index_t idx_N0_M01_local = idx_N0 + idx_M01 * N0;
/**
* idxN0
*
* |< mtx N >|
*
* NPerBlock NPerBlock NPerBlock NPerBlock
* N_0 N_1 N_2 N_3
* - |-----------|-----------|-----------|-----|-----|-
* ^ | - - 0 |/----> 2 | | | |
* | | | / | | | | | M_0 MPerBlock
* | M | /| | | | | |
* |-0---|---/-|-----|-----|-----------|-----|-----|-
* | 1 | / | | | blockid | | |
* idxM0 | | | / | V | 5 | | | M_1 MPerBlock
* | - V 1 | - 3 | | | |
* |-----------|-----------|-----------|-----|-----|-
* mtx M | | | | | |
* | | | | | | M_2 MPerBlock
* | | | | | |
* |-----------|-----------|-----------|-----|-----|-
* | | | | | |
* | | | | | | M_3 MPerBlock
* | | | | | |
* |-----------|-----------|-----------|-----|-----|-
* V | | | | | |
* - |-----------|-----------|-----------|-----|-----|- M_4 MPerBlock
* | | | | | |
* |-----------|-----------|-----------|-----|-----|-
* Example:
* assume:
* M0 = 5
* N0 = 4
* block_1d_id = 5
* M01 = 2
*
* idx_N0 = 1
* idx_M0 = 1
* M01_adapt = 2
* idx_M00 = 0
* idx_M01 = 1
* idx_N0_M01_local = 5
* output {1, 2}
*/
return make_tuple(idx_N0_M01_local % M01_adapt + idx_M00 * M01_,
idx_N0_M01_local / M01_adapt);
}
......@@ -165,11 +236,18 @@ struct BlockToCTileMap_M00_N0_M01Adapt
return true; // always valid provided that user gets grid size from CalculateGridSize()
}
__host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const { return true; }
private:
index_t M_;
index_t N_;
index_t M01_;
CGridDesc_M_N c_grid_desc_m_n_;
};
// keep the redundant type argument for backward compatibility
template <index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N>
struct BlockToCTileMap_M00_N0_M01Adapt : BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>
{
using BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>::
BlockToCTileMap_M00_N0_M01Adapt;
};
// 2D slices of column-vectors in 3D space
......@@ -543,4 +621,52 @@ struct OffsettedBlockToCTileMap
index_t block_start_;
};
/**
* @brief Simple tile mapping which creates 3D grid of block of threads.
*
* @paragraph Description
* This Block-to-C-tile-map creates a 3D grid (n_blocks, m_blocks, z_blocks) of thread
* blocks. The first 2D are regular 2D tiles created by division of output GEMM
* dimenions by corresponding tile size. The third dimension (Z) is a k-split dimension,
* which denotes the number of blocks we use to divide work on GEMM K dimension onto.
*
* @tparam MPerBlock Output block tile size in M dimension.
* @tparam NPerBlock Output block tile size in N dimension.
*/
template <index_t MPerBlock, index_t NPerBlock>
struct BlockToCTileMap_3DGrid_KSplit
{
__host__ __device__ BlockToCTileMap_3DGrid_KSplit() = default;
__host__ __device__ constexpr auto
CalculateGridSize(index_t M, index_t N, index_t k_split) const
{
// Create 3D grid
const auto M0 = math::integer_divide_ceil(M, MPerBlock);
const auto N0 = math::integer_divide_ceil(N, NPerBlock);
return std::make_tuple(N0, M0, k_split);
}
template <typename TopIdx>
__device__ constexpr auto CalculateBottomIndex(const TopIdx&) const
{
return make_tuple(blockIdx.z, blockIdx.y, blockIdx.x);
}
template <typename CTileIdx, typename CTileDim>
__host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
const CTileDim& /* c_tile_dim */) const
{
return true; // always valid provided that user gets grid size from CalculateGridSize()
}
template <typename CGridDesc_M_N>
__host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
{
return true;
}
};
} // namespace ck
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment