Skip to content

GitLab

Unverified Commit 33e78b9a authored by arai713's avatar arai713 Committed by GitHub
Browse files

Merge branch 'develop' into transpose_5d

parents 6bfdd98a 98fd41f5
Show whitespace changes
Inline Side-by-side
Showing with 1903 additions and 400 deletions
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View file @ 33e78b9a
......@@ -462,7 +462,6 @@ struct mfma_type<MfmaInstr::mfma_f64_16x16x4f64>
}
};
#if defined CK_ENABLE_FP8
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16f8f8>
{
......@@ -506,9 +505,7 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32f8f8>
intrin_mfma_f32_16x16x32f8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
#endif
#if defined CK_ENABLE_BF8
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16bf8bf8>
{
......@@ -552,9 +549,7 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32bf8bf8>
intrin_mfma_f32_16x16x32bf8bf8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16f8bf8>
{
......@@ -598,9 +593,7 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32f8bf8>
intrin_mfma_f32_16x16x32f8bf8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16bf8f8>
{
......@@ -644,7 +637,6 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32bf8f8>
intrin_mfma_f32_16x16x32bf8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
#endif
template <typename base_type,
index_t MPerXdlops,
......@@ -792,7 +784,6 @@ struct MfmaSelector
}
#endif
#if defined CK_ENABLE_FP8
template <>
static constexpr auto GetMfma<f8_t, 32, 32>()
{
......@@ -804,9 +795,7 @@ struct MfmaSelector
{
return MfmaInstr::mfma_f32_16x16x32f8f8;
}
#endif
#if defined CK_ENABLE_BF8
template <>
static constexpr auto GetMfma<bf8_t, 32, 32>()
{
......@@ -818,9 +807,7 @@ struct MfmaSelector
{
return MfmaInstr::mfma_f32_16x16x32bf8bf8;
}
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
static constexpr auto GetMfma<f8_t, 32, 32, bf8_t>()
{
......@@ -832,9 +819,7 @@ struct MfmaSelector
{
return MfmaInstr::mfma_f32_16x16x32f8bf8;
}
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
static constexpr auto GetMfma<bf8_t, 32, 32, f8_t>()
{
......@@ -846,7 +831,6 @@ struct MfmaSelector
{
return MfmaInstr::mfma_f32_16x16x32bf8f8;
}
#endif
static constexpr auto selected_mfma =
mfma_type<GetMfma<base_type, MPerXdlops, NPerXdlops, additional_type>()>{};
......@@ -1051,18 +1035,10 @@ struct XdlopsGemm
static_assert(
is_same<base_type, double>::value || is_same<base_type, float>::value ||
is_same<base_type, half_t>::value || is_same<base_type, bhalf_t>::value ||
is_same<base_type, int8_t>::value
#if defined CK_ENABLE_FP8
|| is_same<base_type, f8_t>::value
#endif
#if defined CK_ENABLE_BF8
|| is_same<base_type, bf8_t>::value
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
|| (is_same<base_type, f8_t>::value && is_same<additional_type, bf8_t>::value) ||
(is_same<base_type, bf8_t>::value && is_same<additional_type, f8_t>::value)
#endif
,
is_same<base_type, int8_t>::value || is_same<base_type, f8_t>::value ||
is_same<base_type, bf8_t>::value ||
(is_same<base_type, f8_t>::value && is_same<additional_type, bf8_t>::value) ||
(is_same<base_type, bf8_t>::value && is_same<additional_type, f8_t>::value),
"base base_type must be double, float, half, bfloat16, int8_t, f8_t or bf8_t!");
static_for<0, KPack / mfma_instr.k_per_blk, 1>{}([&](auto k) {
......
include/ck/utility/amd_xdlops.hpp
View file @ 33e78b9a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_AMD_XDLOPS_HPP
#define CK_AMD_XDLOPS_HPP
#include "data_type.hpp"
#pragma once
namespace ck {
......@@ -355,7 +352,6 @@ struct intrin_mfma_f64_16x16x4f64<16, 16>
}
};
#if defined CK_ENABLE_FP8
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16f8f8;
......@@ -418,9 +414,7 @@ struct intrin_mfma_f32_16x16x32f8f8<16, 16>
#endif
}
};
#endif
#if defined CK_ENABLE_BF8
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16bf8bf8;
......@@ -483,9 +477,7 @@ struct intrin_mfma_f32_16x16x32bf8bf8<16, 16>
#endif
}
};
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16f8bf8;
......@@ -548,9 +540,7 @@ struct intrin_mfma_f32_16x16x32f8bf8<16, 16>
#endif
}
};
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16bf8f8;
......@@ -613,6 +603,5 @@ struct intrin_mfma_f32_16x16x32bf8f8<16, 16>
#endif
}
};
#endif
} // namespace ck
#endif
include/ck/utility/data_type.hpp
View file @ 33e78b9a
......@@ -9,15 +9,9 @@ namespace ck {
using bhalf_t = ushort;
using half_t = _Float16;
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
using int4_t = _BitInt(4);
#endif
#if defined CK_ENABLE_FP8
using f8_t = _BitInt(8);
#endif
#if defined CK_ENABLE_BF8
using bf8_t = unsigned _BitInt(8);
#endif
// vector_type
template <typename T, index_t N>
......@@ -148,23 +142,19 @@ struct scalar_type<int4_t>
};
#endif
#if defined CK_ENABLE_FP8
template <>
struct scalar_type<f8_t>
{
using type = f8_t;
static constexpr index_t vector_size = 1;
};
#endif
#if defined CK_ENABLE_BF8
template <>
struct scalar_type<bf8_t>
{
using type = bf8_t;
static constexpr index_t vector_size = 1;
};
#endif
template <typename T>
struct vector_type<T, 1>
......@@ -968,24 +958,20 @@ using int8x32_t = typename vector_type<int8_t, 32>::type;
using int8x64_t = typename vector_type<int8_t, 64>::type;
// f8
#if defined CK_ENABLE_FP8
using f8x2_t = typename vector_type<f8_t, 2>::type;
using f8x4_t = typename vector_type<f8_t, 4>::type;
using f8x8_t = typename vector_type<f8_t, 8>::type;
using f8x16_t = typename vector_type<f8_t, 16>::type;
using f8x32_t = typename vector_type<f8_t, 32>::type;
using f8x64_t = typename vector_type<f8_t, 64>::type;
#endif
// bf8
#if defined CK_ENABLE_BF8
using bf8x2_t = typename vector_type<bf8_t, 2>::type;
using bf8x4_t = typename vector_type<bf8_t, 4>::type;
using bf8x8_t = typename vector_type<bf8_t, 8>::type;
using bf8x16_t = typename vector_type<bf8_t, 16>::type;
using bf8x32_t = typename vector_type<bf8_t, 32>::type;
using bf8x64_t = typename vector_type<bf8_t, 64>::type;
#endif
template <typename T>
struct NumericLimits
......@@ -1033,7 +1019,6 @@ struct NumericLimits<int4_t>
};
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
#if defined CK_ENABLE_FP8
template <>
struct NumericLimits<f8_t>
{
......@@ -1056,9 +1041,7 @@ struct NumericLimits<f8_t>
__host__ __device__ static constexpr f8_t QuietNaN() { return f8_t(binary_qnan); }
};
#endif
#if defined CK_ENABLE_BF8
template <>
struct NumericLimits<bf8_t>
{
......@@ -1081,7 +1064,6 @@ struct NumericLimits<bf8_t>
__host__ __device__ static constexpr bf8_t QuietNaN() { return bf8_t(binary_qnan); }
};
#endif
template <typename T>
struct NumericUtils
......@@ -1093,6 +1075,7 @@ struct NumericUtils<float>
{
static constexpr int exp = 8;
static constexpr int mant = 23;
static constexpr int bias = 127;
static constexpr uint32_t nan_mask = 0x7F800000;
static constexpr uint32_t head_mask = 0xFF800000;
static constexpr uint32_t mant_mask = 0x7FFFFF;
......@@ -1109,6 +1092,7 @@ struct NumericUtils<half_t>
{
static constexpr int exp = 5;
static constexpr int mant = 10;
static constexpr int bias = 15;
static constexpr uint16_t nan_mask = 0x7C00;
static constexpr uint16_t head_mask = 0xFC00;
static constexpr uint16_t mant_mask = 0x3FF;
......@@ -1120,22 +1104,21 @@ struct NumericUtils<half_t>
using bitwise_type = uint16_t;
};
#if defined CK_ENABLE_FP8
template <>
struct NumericUtils<f8_t>
{
static constexpr int exp = 4;
static constexpr int mant = 3;
static constexpr int bias = 8; // negative zero nan mode
// static constexpr int bias = 7; // ieee mode
};
#endif
#if defined CK_ENABLE_BF8
template <>
struct NumericUtils<bf8_t>
{
static constexpr int exp = 5;
static constexpr int mant = 2;
static constexpr int bias = 16; // negative zero nan mode
// static constexpr int bias = 15; // ieee mode
};
#endif
} // namespace ck
include/ck/utility/f8_utils.hpp
View file @ 33e78b9a
......@@ -5,9 +5,6 @@
#include "ck/utility/data_type.hpp"
// these conversions are disabled if native conversions available
#if !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__)
#if defined CK_ENABLE_FP8 || defined CK_ENABLE_BF8
namespace ck {
// fp8 rounding modes
......@@ -19,6 +16,9 @@ enum class f8_rounding_mode
stochastic
};
__host__ inline int clz(uint32_t x) { return __builtin_clz(x); }
__device__ inline int clz(uint32_t x) { return __clz(x); }
} // namespace ck
namespace ck::utils {
......@@ -36,7 +36,7 @@ __host__ __device__ Y run_cast_to_f8(X x, uint32_t rng)
constexpr int in_exp = NumericUtils<X>::exp;
constexpr int in_mant = NumericUtils<X>::mant;
int exponent;
int exponent, bias;
uint32_t head, mantissa, sign;
// nan code is same for float and half
constexpr Y nan_code = 0x80;
......@@ -51,12 +51,11 @@ __host__ __device__ Y run_cast_to_f8(X x, uint32_t rng)
mantissa = x_bitwise & NumericUtils<X>::mant_mask;
exponent = (head >> in_mant) & NumericUtils<X>::exp_mask;
sign = head >> (in_exp + in_mant);
bias = NumericUtils<X>::bias;
uint32_t signed_inf = (sign << (in_exp + in_mant)) + (((1 << in_exp) - 1) << in_mant);
uint32_t drop_mask = (1 << (in_mant - out_mant)) - 1;
constexpr int max_exp = (1 << out_exp) - (negative_zero_nan ? 1 : 2);
constexpr int exp_low_cutoff =
(1 << (in_exp - 1)) - (1 << (out_exp - 1)) + 1 - (negative_zero_nan ? 1 : 0);
if constexpr(negative_zero_nan)
{
......@@ -69,56 +68,107 @@ __host__ __device__ Y run_cast_to_f8(X x, uint32_t rng)
return signed_inf + (mantissa != 0 ? 1 : 0);
}
// if input is half and output is bf8
if((NumericUtils<X>::mant == 10) && (NumericUtils<Y>::mant == 2) && negative_zero_nan &&
exponent == 0)
{
exponent += 1;
while(mantissa < (1 << in_mant))
{
mantissa <<= 1;
exponent -= 1;
}
mantissa &= ~(1 << in_mant);
}
// check if x is 0.0
if(x_bitwise == 0)
return 0;
exponent -= exp_low_cutoff - 1;
if(exponent <= 0)
drop_mask = (1 << (in_mant - out_mant + 1 - exponent)) - 1;
mantissa += 1 << in_mant;
// apply random number if needed
mantissa += (stoch ? rng : mantissa) & drop_mask;
if(mantissa >= (2 << in_mant))
// First need to check if it is normal or denorm as there is a difference of implict 1
// Then need to adjust the exponent to align with the F8 exponent, in the meanwhile, shift
// The mantissa. Then for stochastic rounding, add rng to mantissa and truncate. And for
// RNE, no need to add rng. Then probably need to check whether there is carry and adjust
// exponent and mantissa again3
// For IEEE bias mode, the bias is 2^(k-1)-1 where k is the width of exponent bits
const int out_bias = (1 << (out_exp - 1)) - 1 + (negative_zero_nan ? 1 : 0);
const int out_denormal_act_exponent = 1 - out_bias; // actual exponent of f8 denormal
// act_exponent is the actual exponent of fp32/fp16 (after subtracting bias)
// out_exponent is the converted f8 exponent with bias encoding
// exponent_diff is the diff between fp32/fp16 exponent and f8 exponent,
// the difference needs to be adjusted and mantissa shifted
int act_exponent, out_exponent, exponent_diff;
if(exponent == 0)
{ // fp32/fp16 is in denormal.
/* fp32 denormal is below 2^-127 so it is usually not a concern here, we mostly concern fp16
here. In this case, f8 is usually in denormal. But there could be exceptions. fp16 denormal has
exponent bias 15 while bf8 with NANOO has exponent bias 16. It means that there are some numbers in
fp16 denormal but they are bf8 (NANOO) normals - smallest bf8 (NANOO) normal is 2^-15. fp16 numbers
where exponent==0 (actual exponent -14) and highest bit of mantissa is 1 are bf8 (NANOO) normal.
In this case, the fp16 mantissa should be shift left by 1 */
act_exponent = exponent - bias + 1;
exponent_diff = out_denormal_act_exponent -
act_exponent; // actual exponent is exponent-bias+1 as it is denormal
}
else
{ // fp32/fp16 is normal with implicit 1
act_exponent = exponent - bias;
if(act_exponent <= out_denormal_act_exponent)
{
mantissa >>= 1;
exponent++;
/* This is the case where fp32/fp16 is normal but it is in f8 denormal range.
For example fp8 nanoo mode, denormal exponent is -7, but if the fp32/fp16
actual exponent is -7, it is actually larger due to the implict 1,
Therefore it needs to be adjust to -6 and mantissa shift right by 1.
So for fp32/fp16, exponent -8 is the cut point to convert to fp8 nanoo */
exponent_diff = out_denormal_act_exponent - act_exponent;
}
else
{ // both fp32/fp16 and f8 are in normal range
exponent_diff =
0; // exponent_diff=0 does not mean there is no difference for this case,
// act_exponent could be larger. Just that it does not need shift mantissa
}
mantissa += (1 << in_mant); // Add the implicit 1 into mantissa
}
mantissa >>= (in_mant - out_mant);
// check negative exponent
if(exponent <= 0)
bool midpoint = (mantissa & ((1 << (in_mant - out_mant + exponent_diff)) - 1)) ==
(1 << (in_mant - out_mant + exponent_diff - 1));
/* This part is a bit tricky. The judgment of whether it is a tie needs to be done before we
shift right as shift right could rip off some residual part and make something not midpoint look
like midpoint. For example, the fp16 number 0x1002 (0 00100 0000000010), it is larger than
midpoint, but after shift right by 4 bits, it would look like midpoint. */
if(exponent_diff > 0)
mantissa >>= exponent_diff;
else if(exponent_diff == -1)
mantissa <<= -exponent_diff;
bool implicit_one = mantissa & (1 << in_mant);
// if there is no implict 1, it means the f8 is denormal and need to adjust to denorm exponent
out_exponent =
(act_exponent + exponent_diff) /*actual f8 exponent*/ + out_bias - (implicit_one ? 0 : 1);
// Now we have the exponent and mantissa adjusted
bool odd =
mantissa &
(1 << (in_mant - out_mant)); // if the least significant bit that is not truncated is 1
mantissa += (stoch ? rng : (midpoint ? (odd ? mantissa : mantissa - 1) : mantissa)) & drop_mask;
// Now we deal with overflow
if(out_exponent == 0)
{
if(x_bitwise == 0)
return 0;
if((1 << in_mant) & mantissa)
{
out_exponent = 1; // denormal overflow to become normal, promote exponent
// No need to make 1 implicit now as it will be addressed later
}
}
else
{
// subnormal range; represented by a subnormal float8 (exponent 0)
// and involves loss of accuracy
mantissa >>= 1 - exponent;
exponent = 0;
if((1 << (in_mant + 1)) & mantissa)
{
mantissa >>= 1;
out_exponent++;
// No need to make 1 implicit now as it will be addressed later
}
}
// above range: quantize to maximum possible float of the same sign
else if(exponent > max_exp)
mantissa >>= (in_mant - out_mant);
if(out_exponent > max_exp)
{
if(clip)
{
mantissa = (1 << out_mant) - 1;
exponent = max_exp;
out_exponent = max_exp;
}
else
{
......@@ -127,10 +177,10 @@ __host__ __device__ Y run_cast_to_f8(X x, uint32_t rng)
}
// check if x is 0.0 or -0.0
if(exponent == 0 && mantissa == 0)
if(out_exponent == 0 && mantissa == 0)
return negative_zero_nan ? 0 : (sign << (out_exp + out_mant));
mantissa &= (1 << out_mant) - 1;
return (sign << (out_exp + out_mant)) | (exponent << out_mant) | mantissa;
return (sign << (out_exp + out_mant)) | (out_exponent << out_mant) | mantissa;
}
template <typename X, typename Y, bool negative_zero_nan>
......@@ -196,12 +246,9 @@ __host__ __device__ Y run_cast_from_f8(X x)
if(exponent == 0)
{
// guaranteed mantissa!=0 since cases 0x0 and 0x80 are handled above
exponent++;
while(mantissa < (1 << in_mant))
{
mantissa <<= 1;
exponent--;
}
int sh = 1 + clz(mantissa) - (32 - in_mant);
mantissa <<= sh;
exponent += 1 - sh;
mantissa &= ((1 << in_mant) - 1);
}
exponent += exp_low_cutoff - 1;
......@@ -244,5 +291,3 @@ __host__ __device__ Y cast_from_f8(X x)
}
} // namespace ck::utils
#endif // #if defined CK_ENABLE_FP8 || defined CK_ENABLE_BF8
#endif // #if !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__)
include/ck/utility/inner_product.hpp
View file @ 33e78b9a
......@@ -192,6 +192,8 @@ inner_product<int8x4_t, int8x4_t, int32_t>(const int8x4_t& a, const int8x4_t& b,
#else
c = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b), c, false);
#endif
#elif defined(CK_USE_AMD_V_DOT4_I32_I8_GFX11)
c = __builtin_amdgcn_sudot4(true, bit_cast<int32_t>(a), true, bit_cast<int32_t>(b), c, false);
#else
const vector_type<int8_t, 4> a_vector{a};
const vector_type<int8_t, 4> b_vector{b};
......
include/ck/utility/math.hpp
View file @ 33e78b9a
......@@ -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 @ 33e78b9a
......@@ -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 @ 33e78b9a
......@@ -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 @ 33e78b9a
......@@ -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)
......@@ -344,7 +371,7 @@ inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return f8_convert_sr<f8_t>(type_convert<float>(x));
#elif 0
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
......@@ -353,13 +380,9 @@ inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
return utils::
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#else
return type_convert<f8_t>(type_convert<float>(x));
#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)
......@@ -395,7 +418,7 @@ inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return f8_convert_sr<f8_t>(type_convert<float>(x));
#elif 0
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
......@@ -405,10 +428,7 @@ inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
return utils::
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#else
return type_convert<bf8_t>(type_convert<float>(x));
#endif
}
#endif
} // namespace ck
library/include/ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp
View file @ 33e78b9a
......@@ -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 @ 33e78b9a
......@@ -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 @ 33e78b9a
......@@ -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 @ 33e78b9a
......@@ -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
......@@ -78,8 +85,8 @@ 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);
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) /
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.acc_elementwise_op_(y, y);
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 @ 33e78b9a
......@@ -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 @ 33e78b9a
......@@ -20,8 +20,9 @@ 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
......@@ -36,15 +37,19 @@ 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)
......@@ -55,10 +60,12 @@ 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
......@@ -69,8 +76,8 @@ struct ReferenceLayernorm : public device::BaseOperator
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 +86,7 @@ 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;
}
......@@ -90,17 +97,21 @@ 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 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 * arg.gamma_n_(n)) + arg.beta_n_(n);
arg.acc_elementwise_op_(y_val, y_val);
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;
......@@ -140,13 +151,23 @@ 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)
{
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 @ 33e78b9a
......@@ -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 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 @ 33e78b9a
// 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 @ 33e78b9a
......@@ -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)
{
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_kknn_instance(
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);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
}
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);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance(
}
else if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mnnn_instance(
op_ptrs);
}
}
}
#endif // CK_ENABLE_FP64
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<ADataType, ck::half_t> && is_same_v<BDataType, ck::half_t> &&
is_same_v<EDataType, ck::half_t>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
if constexpr(is_same_v<ComputeDataType, float>)
{
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_kknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_knnn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mknn_instance(
op_ptrs);
add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mnnn_instance(
op_ptrs);
}
}
}
#endif // CK_ENABLE_FP16
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, ck::bhalf_t> &&
is_same_v<EDataType, ck::bhalf_t>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
if constexpr(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 @ 33e78b9a
......@@ -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()
{
......@@ -154,6 +429,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
is_same_v<EDataType, float>)
{
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_f32_f32_f32_kkn_instance(
op_ptrs);
......@@ -164,13 +441,38 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
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)
{
if constexpr(is_same_v<ComputeDataType, double>)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
op_ptrs);
......@@ -181,8 +483,60 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
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 @ 33e78b9a
......@@ -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);
}
}
}
......
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