Skip to content

GitLab

Commit 7572a691 authored by coderfeli's avatar coderfeli
Browse files

merge develop

parents 7796fc73 6b6fcd37
Hide whitespace changes
Inline Side-by-side
Showing with 935 additions and 178 deletions
include/ck_tile/core/numeric/vector_type.hpp
View file @ 7572a691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -200,4 +200,21 @@ using bf8x32_t = bf8_t __attribute((ext_vector_type(32)));
using bf8x64_t = bf8_t __attribute((ext_vector_type(64)));
#endif
CK_TILE_HOST fp16x2_t pk_add_f16(const fp16x2_t& x, const fp16x2_t& y)
{
fp16x2_t vector_res;
vector_res.x = x.x + y.x;
vector_res.y = x.y + y.y;
return vector_res;
}
CK_TILE_DEVICE fp16x2_t pk_add_f16(const fp16x2_t& x, const fp16x2_t& y)
{
fp16x2_t c;
asm volatile("v_pk_add_f16 %0, %1, %2" : "=v"(c) : "v"(x), "v"(y));
return c;
}
} // namespace ck_tile
include/ck_tile/core/tensor/tile_window.hpp
View file @ 7572a691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -18,8 +18,17 @@
namespace ck_tile {
// Note: this tile window do not support single issue
// you need to use tile_window_linear structure for this purpose
/**
* @brief This class provides tile (windowed) view and access to the device memory.
*
* @note This tile window does not support single issue you need to use tile_window_linear
* structure for this purpose
*
* @tparam BottomTensorView_ Class describing & holding device tensor memory.
* @tparam WindowLengths_ Spatial sizes of windowed view on tensor.
* @tparam StaticTileDistribution_ Thread distribution (mapping) into Tile dimensions
* @tparam NumCoord TBD
*/
template <typename BottomTensorView_,
typename WindowLengths_,
typename StaticTileDistribution_,
......@@ -1009,6 +1018,14 @@ CK_TILE_DEVICE void move_tile_window(
window.move(step);
}
/**
* @brief This class provides description of tile windowed view on the device memory.
*
* @note This class does not provide any functions to read or modify device memory.
*
* @tparam BottomTensorView_ Class describing & holding device tensor memory.
* @tparam WindowLengths_ Spatial sizes of windowed view on tensor.
*/
template <typename BottomTensorView_, typename WindowLengths_>
struct tile_window_with_static_lengths
{
......
include/ck_tile/core/tensor/transpose_tile.hpp 0 → 100644
View file @ 7572a691
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core/config.hpp"
#include "ck_tile/core/numeric/integer.hpp"
#include "ck_tile/core/numeric/integral_constant.hpp"
#include "ck_tile/core/utility/functional.hpp"
#include "ck_tile/core/algorithm/coordinate_transform.hpp"
#include "ck_tile/core/algorithm/space_filling_curve.hpp"
#include "ck_tile/core/container/container_helper.hpp"
#include "ck_tile/core/container/thread_buffer.hpp"
#include "ck_tile/core/container/statically_indexed_array.hpp"
#include "ck_tile/core/numeric/math.hpp"
#include "ck_tile/core/utility/type_traits.hpp"
#include "ck_tile/core/tensor/tile_elementwise.hpp"
#include "ck_tile/core/utility/transpose_vectors.hpp"
namespace ck_tile {
namespace detail {
template <typename OutTensor, typename InTensor>
CK_TILE_DEVICE void transpose_tile2d_impl_in_thread(OutTensor& out_tensor,
const InTensor& in_tensor)
{
constexpr auto I0 = number<0>{};
static_assert(std::is_same_v<typename InTensor::DataType, typename OutTensor::DataType>,
"Data type for InTensor and OutTensor must be the same!");
using DataType = typename InTensor::DataType;
constexpr auto y_in_desc = InTensor::get_tile_distribution().get_ys_to_d_descriptor();
constexpr auto y_out_desc = OutTensor::get_tile_distribution().get_ys_to_d_descriptor();
// y_dim_out_to_in
// For swapped Hs tile case I need only get_rh_minor_to_y
// since rh_major are already swapped due to swapped Hs.
constexpr auto get_rh_minor_to_y = [](auto dstr_tensor) {
using DstrEncode = typename decltype(dstr_tensor.get_tile_distribution())::DstrEncode;
map<index_t, index_t> rh_minor_to_y_;
static_for<0, DstrEncode::NDimY, 1>{}([&](auto i) {
constexpr index_t rh_minor = DstrEncode::ys_to_rhs_minor_[i];
rh_minor_to_y_(rh_minor) = i;
});
return rh_minor_to_y_;
};
// In swapped Hs case <Y,X> -> <X,Y> tile
// we have same rh_major, but reversed rh_minor!
constexpr auto rh_minor_to_y_in = get_rh_minor_to_y(InTensor{});
constexpr auto rh_minor_to_y_out = get_rh_minor_to_y(OutTensor{});
// Is this really needed?? Should we have simple reverse here??
constexpr auto y_dim_out_to_in = [&] {
map<index_t, index_t> y_dim_out_to_in_;
for(const auto& [rh_minor, y_out] : rh_minor_to_y_out)
{
y_dim_out_to_in_(y_out) = rh_minor_to_y_in[rh_minor];
}
return y_dim_out_to_in_;
}();
constexpr index_t NDimY = InTensor::get_tile_distribution().get_num_of_dimension_y();
constexpr auto y_lengths = to_sequence(y_in_desc.get_lengths());
// input and output vector dim in the order of input Y dims
constexpr index_t y_dim_vec_in = NDimY - 1;
constexpr index_t y_dim_vec_out = y_dim_out_to_in[NDimY - 1];
// vector lengths
constexpr index_t vec_length_in = y_lengths[y_dim_vec_in];
constexpr index_t vec_length_out = y_lengths[y_dim_vec_out];
// # of vectors
constexpr index_t num_vec_in = vec_length_out;
constexpr index_t num_vec_out = vec_length_in;
using InVec = array<DataType, vec_length_in>;
using OutVec = array<DataType, vec_length_out>;
// SFC
constexpr auto scalars_per_access_arr = generate_array(
[&](auto i) { return (i == y_dim_vec_in or i == y_dim_vec_out) ? y_lengths[i] : 1; },
number<NDimY>{});
constexpr auto scalars_per_access = TO_SEQUENCE(scalars_per_access_arr, NDimY);
using SFC_Y = space_filling_curve<decltype(y_lengths),
typename arithmetic_sequence_gen<0, NDimY, 1>::type,
decltype(scalars_per_access)>;
constexpr index_t num_access = SFC_Y::get_num_of_access();
static_assert(num_access > 0, "wrong! num_access should be larger than 0");
// in/out vectors to be transposed
thread_buffer<InVec, num_vec_in> in_vectors;
thread_buffer<OutVec, num_vec_out> out_vectors;
// loop over SFC and do transpose
static_for<0, num_access, 1>{}([&](auto iAccess) {
// data index [y0, y1, ...] in the order of input tensor
constexpr auto idx_y_start = SFC_Y::get_index(iAccess);
// get input vectors
static_for<0, num_vec_in, 1>{}([&](auto i) {
constexpr auto idx_y_in = generate_tuple(
[&](auto ii) {
return ii == y_dim_vec_out ? idx_y_start[ii] + i : idx_y_start[ii];
},
number<NDimY>{});
constexpr index_t in_offset = y_in_desc.calculate_offset(idx_y_in);
static_assert(in_offset % vec_length_in == 0);
in_vectors(i).template get_as<InVec>()(I0) =
in_tensor.get_thread_buffer()
.template get_as<InVec>()[number<in_offset / vec_length_in>{}];
});
// transpose
transpose_vectors<DataType, num_vec_in, num_vec_out>{}(in_vectors, out_vectors);
// set output vectors
static_for<0, num_vec_out, 1>{}([&](auto i) {
constexpr auto idx_y_out_tmp = generate_array(
[&](auto ii) { return ii == y_dim_vec_in ? idx_y_start[ii] + i : idx_y_start[ii]; },
number<NDimY>{});
constexpr auto idx_y_out =
container_reorder_given_new2old(idx_y_out_tmp, y_dim_out_to_in);
constexpr index_t out_offset = y_out_desc.calculate_offset(idx_y_out);
static_assert(out_offset % vec_length_out == 0);
out_tensor.get_thread_buffer().template set_as<OutVec>(
number<out_offset / vec_length_out>{},
out_vectors[i].template get_as<OutVec>()[I0]);
});
});
}
} // namespace detail
template <typename OutTensor, typename InTensor>
CK_TILE_DEVICE void transpose_tile2d(OutTensor& out, const InTensor& in)
{
using InDataType = typename InTensor::DataType;
using OutDataType = typename OutTensor::DataType;
using InTileDistr = typename InTensor::StaticTileDistribution;
using OutTileDistr = typename OutTensor::StaticTileDistribution;
using InDstrEncode = typename InTileDistr::DstrEncode;
using OutDstrEncode = typename OutTileDistr::DstrEncode;
using InThreadTensorDesc = typename InTensor::ThreadTensorDesc;
using OutThreadTensorDesc = typename OutTensor::ThreadTensorDesc;
// Ys:
constexpr auto in_thread_desc_lengths = InThreadTensorDesc{}.get_lengths();
constexpr auto out_thread_desc_lengths = OutThreadTensorDesc{}.get_lengths();
// type convert
const auto in_tmp = [&]() {
if constexpr(std::is_same_v<OutDataType, InDataType>)
{
return in;
}
else
{
return tile_elementwise_in(type_convert<OutDataType, InDataType>, in);
}
}();
// Scenario where we switch from tile <Y, X> -> <X, Y> - only 2D tiles!
// we preserve Ps but swap Ys: <Y1, Y0> -> <Y0, Y1>
if constexpr(InDstrEncode::rs_lengths_ == OutDstrEncode::rs_lengths_ &&
InDstrEncode::hs_lengthss_ == tuple_reverse(OutDstrEncode::hs_lengthss_) &&
InDstrEncode::NDimY == OutDstrEncode::NDimY && InDstrEncode::NDimY == 2 &&
in_thread_desc_lengths == tuple_reverse(out_thread_desc_lengths))
// Any condition on Ps ??
// InDstrEncode::ps_to_rhss_major_ == OutDstrEncode::ps_to_rhss_major_ &&
// InDstrEncode::ps_to_rhss_minor_ == OutDstrEncode::ps_to_rhss_minor_ &&
{
detail::transpose_tile2d_impl_in_thread(out, in_tmp);
}
else
{
static_assert(false, "Provided tensors could not be transposed!");
}
}
} // namespace ck_tile
include/ck_tile/core/utility/amd_address_space.hpp deleted 100644 → 0
View file @ 7796fc73
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core/config.hpp"
// Address Space for AMDGCN
// https://llvm.org/docs/AMDGPUUsage.html#address-space
namespace ck_tile {
#define CK_CONSTANT_ADDRESS_SPACE __attribute__((address_space(4)))
template <typename T>
__device__ T* cast_pointer_to_generic_address_space(T CK_CONSTANT_ADDRESS_SPACE* p)
{
// cast a pointer in "Constant" address space (4) to "Generic" address space (0)
// only c-style pointer cast seems be able to be compiled
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wold-style-cast"
return (T*)p; // NOLINT(old-style-cast)
#pragma clang diagnostic pop
}
template <typename T>
__host__ __device__ T CK_CONSTANT_ADDRESS_SPACE* cast_pointer_to_constant_address_space(T* p)
{
// cast a pointer in "Generic" address space (0) to "Constant" address space (4)
// only c-style pointer cast seems be able to be compiled
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wold-style-cast"
return (T CK_CONSTANT_ADDRESS_SPACE*)p; // NOLINT(old-style-cast)
#pragma clang diagnostic pop
}
} // namespace ck_tile
include/ck_tile/core/utility/transpose_vectors.hpp
View file @ 7572a691
......@@ -68,52 +68,82 @@ struct transpose_vectors
}
else if constexpr(sizeof(S) == 1)
{
static_assert((NX % 4 == 0 && NY % 4 == 0), "wrong!");
static_assert(((NX % 4 == 0 && NY % 4 == 0) || (NX % 2 == 0 && NY % 2 == 0)), "wrong!");
using S4 = array<S, 4>; // typename array<S, 4>::type;
// loop over 4x4 tile and transpose data from vx_tuple into vy_tuple
static_for<0, NY, 4>{}([&](auto iy) {
static_for<0, NX, 4>{}([&](auto ix) {
// 4 int8x4 data from vx_tuple
const int32_t x_s4_0 =
bit_cast<int32_t>(vx_tuple[ix].template get_as<S4>()[iy / I4]);
const int32_t x_s4_1 =
bit_cast<int32_t>(vx_tuple[ix + I1].template get_as<S4>()[iy / I4]);
const int32_t x_s4_2 =
bit_cast<int32_t>(vx_tuple[ix + I2].template get_as<S4>()[iy / I4]);
const int32_t x_s4_3 =
bit_cast<int32_t>(vx_tuple[ix + I3].template get_as<S4>()[iy / I4]);
// transpose
int32_t t_s4_0, t_s4_1;
int32_t y_s4_0, y_s4_1, y_s4_2, y_s4_3;
constexpr int32_t m0 = 0x05010400;
constexpr int32_t m1 = 0x05040100;
constexpr int32_t m2 = 0x07060302;
constexpr int32_t m3 = 0x07030602;
// ex: v_perm_b32(0x 11 22 33 44, 0x 55 66 77 88, 0x 05 01 04 00) -> 0x33774488
// -- -- -- -- -- -- -- -- - - - -
// index 7 6 5 4 3 2 1 0 33 77 44 88
// index is reversed because of little endianness (least significant bits first)
t_s4_0 = __builtin_amdgcn_perm(x_s4_1, x_s4_0, m0);
t_s4_1 = __builtin_amdgcn_perm(x_s4_3, x_s4_2, m0);
y_s4_0 = __builtin_amdgcn_perm(t_s4_1, t_s4_0, m1);
y_s4_1 = __builtin_amdgcn_perm(t_s4_1, t_s4_0, m2);
t_s4_0 = __builtin_amdgcn_perm(x_s4_1, x_s4_0, m3);
t_s4_1 = __builtin_amdgcn_perm(x_s4_3, x_s4_2, m3);
y_s4_2 = __builtin_amdgcn_perm(t_s4_1, t_s4_0, m1);
y_s4_3 = __builtin_amdgcn_perm(t_s4_1, t_s4_0, m2);
// 4 int8x4 data from vy_tuple
vy_tuple(iy).template get_as<S4>()(ix / I4) = bit_cast<S4>(y_s4_0);
vy_tuple(iy + I1).template get_as<S4>()(ix / I4) = bit_cast<S4>(y_s4_1);
vy_tuple(iy + I2).template get_as<S4>()(ix / I4) = bit_cast<S4>(y_s4_2);
vy_tuple(iy + I3).template get_as<S4>()(ix / I4) = bit_cast<S4>(y_s4_3);
using S2 = array<S, 2>; // typename array<S, 4>::type;
if constexpr(NX % 4 == 0 && NY % 4 == 0)
{
// loop over 4x4 tile and transpose data from vx_tuple into vy_tuple
static_for<0, NY, 4>{}([&](auto iy) {
static_for<0, NX, 4>{}([&](auto ix) {
// 4 int8x4 data from vx_tuple
const int32_t x_s4_0 =
bit_cast<int32_t>(vx_tuple[ix].template get_as<S4>()[iy / I4]);
const int32_t x_s4_1 =
bit_cast<int32_t>(vx_tuple[ix + I1].template get_as<S4>()[iy / I4]);
const int32_t x_s4_2 =
bit_cast<int32_t>(vx_tuple[ix + I2].template get_as<S4>()[iy / I4]);
const int32_t x_s4_3 =
bit_cast<int32_t>(vx_tuple[ix + I3].template get_as<S4>()[iy / I4]);
// transpose
int32_t t_s4_0, t_s4_1;
int32_t y_s4_0, y_s4_1, y_s4_2, y_s4_3;
constexpr int32_t m0 = 0x05010400;
constexpr int32_t m1 = 0x05040100;
constexpr int32_t m2 = 0x07060302;
constexpr int32_t m3 = 0x07030602;
// ex: v_perm_b32(0x 11 22 33 44, 0x 55 66 77 88, 0x 05 01 04 00) ->
// 0x33774488
// -- -- -- -- -- -- -- -- - - - -
// index 7 6 5 4 3 2 1 0 33 77 44 88
// index is reversed because of little endianness (least significant bits
// first)
t_s4_0 = __builtin_amdgcn_perm(x_s4_1, x_s4_0, m0);
t_s4_1 = __builtin_amdgcn_perm(x_s4_3, x_s4_2, m0);
y_s4_0 = __builtin_amdgcn_perm(t_s4_1, t_s4_0, m1);
y_s4_1 = __builtin_amdgcn_perm(t_s4_1, t_s4_0, m2);
t_s4_0 = __builtin_amdgcn_perm(x_s4_1, x_s4_0, m3);
t_s4_1 = __builtin_amdgcn_perm(x_s4_3, x_s4_2, m3);
y_s4_2 = __builtin_amdgcn_perm(t_s4_1, t_s4_0, m1);
y_s4_3 = __builtin_amdgcn_perm(t_s4_1, t_s4_0, m2);
// 4 int8x4 data from vy_tuple
vy_tuple(iy).template get_as<S4>()(ix / I4) = bit_cast<S4>(y_s4_0);
vy_tuple(iy + I1).template get_as<S4>()(ix / I4) = bit_cast<S4>(y_s4_1);
vy_tuple(iy + I2).template get_as<S4>()(ix / I4) = bit_cast<S4>(y_s4_2);
vy_tuple(iy + I3).template get_as<S4>()(ix / I4) = bit_cast<S4>(y_s4_3);
});
});
});
}
else if constexpr(NX % 2 == 0 && NY % 2 == 0)
{
static_for<0, NY, 2>{}([&](auto ix) {
static_for<0, NX, 2>{}([&](auto iy) {
const int16_t x_s2_0 =
bit_cast<int16_t>(vx_tuple[ix].template get_as<S2>()[iy / I2]);
const int16_t x_s2_1 =
bit_cast<int16_t>(vx_tuple[ix + I1].template get_as<S2>()[iy / I2]);
constexpr int32_t m0 = 0x05040100;
constexpr int32_t m1 = 0x07060302;
const int32_t x0_32 = static_cast<int32_t>(x_s2_0 & 0xFFFF);
const int32_t x1_32 = static_cast<int32_t>(x_s2_1 & 0xFFFF);
const int32_t y_s2_0 = __builtin_amdgcn_perm(x1_32, x0_32, m0);
const int32_t y_s2_1 = __builtin_amdgcn_perm(x1_32, x0_32, m1);
vy_tuple(iy).template get_as<S2>()[ix / I2] =
bit_cast<S2>(static_cast<int16_t>(y_s2_0 & 0xFFFF));
vy_tuple(iy + I1).template get_as<S2>()[ix / I2] =
bit_cast<S2>(static_cast<int16_t>(y_s2_1 & 0xFFFF));
});
});
}
}
else
{
......
include/ck_tile/core/utility/type_traits.hpp
View file @ 7572a691
......@@ -109,4 +109,22 @@ CK_TILE_HOST_DEVICE PY c_style_pointer_cast(PX p_x)
#pragma clang diagnostic pop
}
template <typename CompareTo, typename... Rest>
struct is_any_of : std::false_type
{
};
template <typename CompareTo, typename FirstType>
struct is_any_of<CompareTo, FirstType> : std::is_same<CompareTo, FirstType>
{
};
template <typename CompareTo, typename FirstType, typename... Rest>
struct is_any_of<CompareTo, FirstType, Rest...>
: std::integral_constant<bool,
std::is_same<CompareTo, FirstType>::value ||
is_any_of<CompareTo, Rest...>::value>
{
};
} // namespace ck_tile
include/ck_tile/core/utility/unary_element_function.hpp
View file @ 7572a691
......@@ -51,16 +51,18 @@ struct composes<F>
template <typename... Ts>
__host__ __device__ composes(Ts&&...)->composes<remove_cvref_t<Ts>...>;
template <typename To>
template <typename SaturateType>
struct saturates
{
template <typename From>
CK_TILE_HOST_DEVICE constexpr auto operator()(const From& from) const
-> std::enable_if_t<std::is_arithmetic_v<From>, From>
// NOTE: this function does not return SaturateType value
// it is user's responsiblity to do further cast or not
template <typename AccType>
CK_TILE_HOST_DEVICE constexpr auto operator()(const AccType& a_) const
-> std::enable_if_t<std::is_arithmetic_v<AccType>, AccType>
{
return clamp(from,
type_convert<From>(numeric<To>::lowest()),
type_convert<From>(numeric<To>::max()));
return clamp(a_,
type_convert<AccType>(numeric<SaturateType>::lowest()),
type_convert<AccType>(numeric<SaturateType>::max()));
}
};
......
include/ck_tile/host.hpp
View file @ 7572a691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/host/arg_parser.hpp"
#include "ck_tile/host/check_err.hpp"
#include "ck_tile/host/concat.hpp"
#include "ck_tile/host/convolution_host_tensor_descriptor_helper.hpp"
#include "ck_tile/host/convolution_parameter.hpp"
#include "ck_tile/host/device_memory.hpp"
......@@ -20,6 +21,7 @@
#include "ck_tile/host/reference/reference_batched_masking.hpp"
#include "ck_tile/host/reference/reference_batched_rotary_position_embedding.hpp"
#include "ck_tile/host/reference/reference_batched_softmax.hpp"
#include "ck_tile/host/reference/reference_batched_transpose.hpp"
#include "ck_tile/host/reference/reference_elementwise.hpp"
#include "ck_tile/host/reference/reference_fused_moe.hpp"
#include "ck_tile/host/reference/reference_gemm.hpp"
......
include/ck_tile/host/check_err.hpp
View file @ 7572a691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -18,6 +18,114 @@
namespace ck_tile {
template <typename ComputeDataType, typename OutDataType, typename AccDataType = ComputeDataType>
double get_relative_threshold(const int number_of_accumulations = 1)
{
using F8 = ck_tile::fp8_t;
using BF8 = ck_tile::bf8_t;
using F16 = ck_tile::half_t;
using BF16 = ck_tile::bf16_t;
using F32 = float;
using I8 = int8_t;
using I32 = int32_t;
static_assert(is_any_of<ComputeDataType, F8, BF8, F16, BF16, F32, I8, I32, int>::value,
"Warning: Unhandled ComputeDataType for setting up the relative threshold!");
double compute_error = 0;
if constexpr(is_any_of<ComputeDataType, I8, I32, int>::value)
{
return 0;
}
else
{
compute_error = std::pow(2, -numeric_traits<ComputeDataType>::mant) * 0.5;
}
static_assert(is_any_of<OutDataType, F8, BF8, F16, BF16, F32, I8, I32, int>::value,
"Warning: Unhandled OutDataType for setting up the relative threshold!");
double output_error = 0;
if constexpr(is_any_of<OutDataType, I8, I32, int>::value)
{
return 0;
}
else
{
output_error = std::pow(2, -numeric_traits<OutDataType>::mant) * 0.5;
}
double midway_error = std::max(compute_error, output_error);
static_assert(is_any_of<AccDataType, F8, BF8, F16, BF16, F32, I8, I32, int>::value,
"Warning: Unhandled AccDataType for setting up the relative threshold!");
double acc_error = 0;
if constexpr(is_any_of<AccDataType, I8, I32, int>::value)
{
return 0;
}
else
{
acc_error = std::pow(2, -numeric_traits<AccDataType>::mant) * 0.5 * number_of_accumulations;
}
return std::max(acc_error, midway_error);
}
template <typename ComputeDataType, typename OutDataType, typename AccDataType = ComputeDataType>
double get_absolute_threshold(const double max_possible_num, const int number_of_accumulations = 1)
{
using F8 = ck_tile::fp8_t;
using BF8 = ck_tile::bf8_t;
using F16 = ck_tile::half_t;
using BF16 = ck_tile::bf16_t;
using F32 = float;
using I8 = int8_t;
using I32 = int32_t;
static_assert(is_any_of<ComputeDataType, F8, BF8, F16, BF16, F32, I8, I32, int>::value,
"Warning: Unhandled ComputeDataType for setting up the absolute threshold!");
auto expo = std::log2(std::abs(max_possible_num));
double compute_error = 0;
if constexpr(is_any_of<ComputeDataType, I8, I32, int>::value)
{
return 0;
}
else
{
compute_error = std::pow(2, expo - numeric_traits<ComputeDataType>::mant) * 0.5;
}
static_assert(is_any_of<OutDataType, F8, BF8, F16, BF16, F32, I8, I32, int>::value,
"Warning: Unhandled OutDataType for setting up the absolute threshold!");
double output_error = 0;
if constexpr(is_any_of<OutDataType, I8, I32, int>::value)
{
return 0;
}
else
{
output_error = std::pow(2, expo - numeric_traits<OutDataType>::mant) * 0.5;
}
double midway_error = std::max(compute_error, output_error);
static_assert(is_any_of<AccDataType, F8, BF8, F16, BF16, F32, I8, I32, int>::value,
"Warning: Unhandled AccDataType for setting up the absolute threshold!");
double acc_error = 0;
if constexpr(is_any_of<AccDataType, I8, I32, int>::value)
{
return 0;
}
else
{
acc_error =
std::pow(2, expo - numeric_traits<AccDataType>::mant) * 0.5 * number_of_accumulations;
}
return std::max(acc_error, midway_error);
}
template <typename T>
std::ostream& operator<<(std::ostream& os, const std::vector<T>& v)
{
......@@ -337,7 +445,11 @@ std::enable_if_t<(std::is_same_v<ranges::range_value_t<Range>, ranges::range_val
}
if(!res)
{
std::cerr << std::setw(12) << std::setprecision(7) << "max err: " << max_err << std::endl;
const float error_percent =
static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
std::cerr << "max err: " << max_err;
std::cerr << ", number of errors: " << err_count;
std::cerr << ", " << error_percent << "% wrong values" << std::endl;
}
return res;
}
......@@ -391,7 +503,11 @@ std::enable_if_t<(std::is_same_v<ranges::range_value_t<Range>, ranges::range_val
}
if(!res)
{
std::cerr << std::setw(12) << std::setprecision(7) << "max err: " << max_err << std::endl;
const float error_percent =
static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
std::cerr << "max err: " << max_err;
std::cerr << ", number of errors: " << err_count;
std::cerr << ", " << error_percent << "% wrong values" << std::endl;
}
return res;
}
......
include/ck_tile/host/concat.hpp 0 → 100644
View file @ 7572a691
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
namespace ck_tile {
template <typename T>
struct IsCharArray : std::false_type
{
};
template <std::size_t N>
struct IsCharArray<char[N]> : std::true_type
{
};
template <std::size_t N>
struct IsCharArray<const char[N]> : std::true_type
{
};
template <std::size_t N>
struct IsCharArray<char (&)[N]> : std::true_type
{
};
template <std::size_t N>
struct IsCharArray<const char (&)[N]> : std::true_type
{
};
template <typename... Ts>
inline constexpr bool AllConvertibleToStringView = ((std::is_convertible_v<Ts, std::string_view> ||
IsCharArray<Ts>::value ||
std::is_same_v<Ts, char>)&&...);
template <typename... Ts>
[[nodiscard]] auto concat(const Ts&... xs)
-> std::enable_if_t<!AllConvertibleToStringView<Ts...>, std::string>
{
using ::operator<<;
thread_local std::ostringstream oss;
oss.str("");
(oss << ... << xs);
return oss.str();
}
template <std::size_t N>
[[nodiscard]] constexpr inline std::size_t getSize(char (&)[N]) noexcept
{
return N;
}
template <std::size_t N>
[[nodiscard]] constexpr inline std::size_t getSize(const char (&)[N]) noexcept
{
return N;
}
[[nodiscard]] constexpr inline std::size_t getSize(const char* s) noexcept
{
const char* end = s;
while(*end++ != 0) {}
return end - s - 1;
}
[[nodiscard]] constexpr inline std::size_t getSize(const char&) noexcept { return 1; }
[[nodiscard]] inline std::size_t getSize(const std::string& s) noexcept { return s.size(); }
[[nodiscard]] constexpr inline std::size_t getSize(const std::string_view& s) noexcept
{
return s.size();
}
template <typename... Ts>
auto concatInto(std::string& result, const Ts&... xs)
-> std::enable_if_t<AllConvertibleToStringView<Ts...>, void>
{
const std::size_t space = (1 + ... + getSize(xs));
result.reserve(result.size() + space);
((result += xs), ...);
}
template <typename... Ts>
[[nodiscard]] auto concat(const Ts&... xs)
-> std::enable_if_t<AllConvertibleToStringView<Ts...>, std::string>
{
std::string result;
concatInto(result, xs...);
return result;
}
// Function for types convertible to std::string_view
template <typename Sep, typename First, typename... Rest>
[[nodiscard]] auto concat(Sep sep, const First& first, const Rest&... rest)
-> std::enable_if_t<AllConvertibleToStringView<First, Rest...>, std::string>
{
std::string result;
result += first;
((result += sep, result += rest), ...);
return result;
}
// Function for other types
template <typename Sep, typename First, typename... Rest>
[[nodiscard]] auto concat(Sep sep, const First& first, const Rest&... rest)
-> std::enable_if_t<!AllConvertibleToStringView<First, Rest...>, std::string>
{
using ::operator<<;
thread_local std::ostringstream oss;
oss.str("");
oss << first;
((oss << sep << rest), ...);
return oss.str();
}
} // namespace ck_tile
include/ck_tile/host/convolution_host_tensor_descriptor_helper.hpp
View file @ 7572a691
......@@ -14,57 +14,41 @@ namespace detail {
template <typename OldLayout>
CK_TILE_HOST std::vector<std::size_t> get_layout_transpose_gnchw_to_old()
{
if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNCW> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKCX> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNKW>)
using namespace ck_tile::tensor_layout::convolution;
if constexpr(is_any_of<OldLayout, GNCW, GKCX, GNKW>::value)
{
return {0, 1, 2, 3};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNCHW> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKCYX> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNKHW>)
else if constexpr(is_any_of<OldLayout, GNCHW, GKCYX, GNKHW>::value)
{
return {0, 1, 2, 3, 4};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNCDHW> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKCZYX> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNKDHW>)
else if constexpr(is_any_of<OldLayout, GNCDHW, GKCZYX, GNKDHW>::value)
{
return {0, 1, 2, 3, 4, 5};
}
if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNWC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKXC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNWK>)
if constexpr(is_any_of<OldLayout, GNWC, GKXC, GNWK>::value)
{
return {0, 1, 3, 2};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNHWC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKYXC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNHWK>)
else if constexpr(is_any_of<OldLayout, GNHWC, GKYXC, GNHWK>::value)
{
return {0, 1, 4, 2, 3};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNDHWC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKZYXC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNDHWK>)
else if constexpr(is_any_of<OldLayout, GNDHWC, GKZYXC, GNDHWK>::value)
{
return {0, 1, 5, 2, 3, 4};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NWGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::KXGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NWGK>)
else if constexpr(is_any_of<OldLayout, NWGC, KXGC, NWGK>::value)
{
return {2, 0, 3, 1};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NHWGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::KYXGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NHWGK>)
else if constexpr(is_any_of<OldLayout, NHWGC, KYXGC, NHWGK>::value)
{
return {3, 0, 4, 1, 2};
}
else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NDHWGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::KZYXGC> ||
std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NDHWGK>)
else if constexpr(is_any_of<OldLayout, NDHWGC, KZYXGC, NDHWGK>::value)
{
return {4, 0, 5, 1, 2, 3};
}
......@@ -83,11 +67,11 @@ template <typename InLayout>
CK_TILE_HOST HostTensorDescriptor
make_input_host_tensor_descriptor_g_n_c_wis_packed(const ck_tile::conv::ConvParam& param)
{
using namespace ck_tile::tensor_layout::convolution;
std::vector<std::size_t> physical_lengths;
if constexpr(std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNCW> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNCHW> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNCDHW>)
if constexpr(is_any_of<InLayout, GNCW, GNCHW, GNCDHW>::value)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.N_),
......@@ -97,9 +81,7 @@ make_input_host_tensor_descriptor_g_n_c_wis_packed(const ck_tile::conv::ConvPara
param.input_spatial_lengths_.begin(),
param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNWC> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNHWC> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNDHWC>)
else if constexpr(is_any_of<InLayout, GNWC, GNHWC, GNDHWC>::value)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.N_),
......@@ -109,9 +91,7 @@ make_input_host_tensor_descriptor_g_n_c_wis_packed(const ck_tile::conv::ConvPara
param.input_spatial_lengths_.begin(),
param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::NWGC> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::NHWGC> ||
std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::NDHWGC>)
else if constexpr(is_any_of<InLayout, NWGC, NHWGC, NDHWGC>::value)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.N_),
static_cast<std::size_t>(param.G_),
......@@ -139,11 +119,11 @@ template <typename WeiLayout>
CK_TILE_HOST HostTensorDescriptor
make_weight_host_tensor_descriptor_g_k_c_xs_packed(const ck_tile::conv::ConvParam& param)
{
using namespace ck_tile::tensor_layout::convolution;
std::vector<std::size_t> physical_lengths;
if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KXC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KYXC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KZYXC>)
if constexpr(is_any_of<WeiLayout, KXC, KYXC, KZYXC>::value)
{
if(param.G_ != 1)
{
......@@ -157,9 +137,7 @@ make_weight_host_tensor_descriptor_g_k_c_xs_packed(const ck_tile::conv::ConvPara
param.filter_spatial_lengths_.begin(),
param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKCX> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKCYX> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKCZYX>)
else if constexpr(is_any_of<WeiLayout, GKCX, GKCYX, GKCZYX>::value)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.K_),
......@@ -169,9 +147,7 @@ make_weight_host_tensor_descriptor_g_k_c_xs_packed(const ck_tile::conv::ConvPara
param.filter_spatial_lengths_.begin(),
param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKXC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKYXC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKZYXC>)
else if constexpr(is_any_of<WeiLayout, GKXC, GKYXC, GKZYXC>::value)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.K_),
......@@ -181,9 +157,7 @@ make_weight_host_tensor_descriptor_g_k_c_xs_packed(const ck_tile::conv::ConvPara
param.filter_spatial_lengths_.begin(),
param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KXGC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KYXGC> ||
std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KZYXGC>)
else if constexpr(is_any_of<WeiLayout, KXGC, KYXGC, KZYXGC>::value)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.K_),
static_cast<std::size_t>(param.G_),
......@@ -211,11 +185,11 @@ template <typename OutLayout>
CK_TILE_HOST HostTensorDescriptor
make_output_host_tensor_descriptor_g_n_k_wos_packed(const ck_tile::conv::ConvParam& param)
{
using namespace ck_tile::tensor_layout::convolution;
std::vector<std::size_t> physical_lengths;
if constexpr(std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNKW> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNKHW> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNKDHW>)
if constexpr(is_any_of<OutLayout, GNKW, GNKHW, GNKDHW>::value)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.N_),
......@@ -226,9 +200,7 @@ make_output_host_tensor_descriptor_g_n_k_wos_packed(const ck_tile::conv::ConvPar
param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
}
// separate from legacy code above
else if constexpr(std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNWK> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNHWK> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNDHWK>)
else if constexpr(is_any_of<OutLayout, GNWK, GNHWK, GNDHWK>::value)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
static_cast<std::size_t>(param.N_),
......@@ -238,9 +210,7 @@ make_output_host_tensor_descriptor_g_n_k_wos_packed(const ck_tile::conv::ConvPar
param.output_spatial_lengths_.begin(),
param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
}
else if constexpr(std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::NWGK> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::NHWGK> ||
std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::NDHWGK>)
else if constexpr(is_any_of<OutLayout, NWGK, NHWGK, NDHWGK>::value)
{
physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.N_),
static_cast<std::size_t>(param.G_),
......
include/ck_tile/host/host_tensor.hpp
View file @ 7572a691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -678,4 +678,43 @@ struct HostTensor
Descriptor mDesc;
Data mData;
};
template <bool is_row_major>
auto host_tensor_descriptor(std::size_t row,
std::size_t col,
std::size_t stride,
bool_constant<is_row_major>)
{
using namespace ck_tile::literals;
if constexpr(is_row_major)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
}
template <bool is_row_major>
auto get_default_stride(std::size_t row,
std::size_t col,
std::size_t stride,
bool_constant<is_row_major>)
{
if(stride == 0)
{
if constexpr(is_row_major)
{
return col;
}
else
{
return row;
}
}
else
return stride;
}
} // namespace ck_tile
include/ck_tile/host/reference/reference_batched_transpose.hpp 0 → 100644
View file @ 7572a691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
#include <thread>
namespace ck_tile {
template <typename Type>
CK_TILE_HOST void reference_batched_transpose(const HostTensor<Type>& x,
HostTensor<Type>& y,
std::string layout_in = "NCHW",
std::string layout_out = "NHWC")
{
const int N = x.mDesc.get_lengths()[0];
auto f = [&](auto batch) {
if(layout_in == "NCHW" && layout_out == "NHWC")
{
const int C = x.mDesc.get_lengths()[1];
const int H = x.mDesc.get_lengths()[2];
const int W = x.mDesc.get_lengths()[3];
for(int c = 0; c < C; ++c)
{
for(int h = 0; h < H; ++h)
{
for(int w = 0; w < W; ++w)
{
Type v_x = x(batch, c, h, w);
y(batch, h, w, c) = v_x;
}
}
}
}
else if(layout_in == "NHWC" && layout_out == "NCHW")
{
const int H = x.mDesc.get_lengths()[1];
const int W = x.mDesc.get_lengths()[2];
const int C = x.mDesc.get_lengths()[3];
for(int h = 0; h < H; ++h)
{
for(int w = 0; w < W; ++w)
{
for(int c = 0; c < C; ++c)
{
Type v_x = x(batch, h, w, c);
y(batch, c, h, w) = v_x;
}
}
}
}
};
make_ParallelTensorFunctor(f, N)(std::thread::hardware_concurrency());
}
} // namespace ck_tile
include/ck_tile/host/reference/reference_fused_moe.hpp
View file @ 7572a691
......@@ -73,7 +73,7 @@ void reference_fused_moe(
ck_tile::index_t tokens,
ck_tile::index_t experts,
ck_tile::index_t hidden_size,
ck_tile::index_t intermediate_size, // this size is for gate/up
ck_tile::index_t intermediate_size, // this size is for gate/up/down
ck_tile::index_t topk,
ck_tile::index_t gate_only)
{
......@@ -82,19 +82,8 @@ void reference_fused_moe(
assert(sorted_expert_ids_host.get_num_of_dimension() == 1);
assert(num_sorted_tiles_host.get_element_size() == 1);
ck_tile::index_t num_sorted_tiles = num_sorted_tiles_host.mData[0] / block_m;
ck_tile::index_t intermediate_size_0 = intermediate_size;
ck_tile::index_t intermediate_size_1 = intermediate_size / (gate_only ? 1 : 2);
// TODO: better remove this in the future, or modify the token_id value
auto get_topk_id = [&](ck_tile::index_t token_id_, ck_tile::index_t expert_id_) {
for(ck_tile::index_t i_ = 0; i_ < topk; i_++)
{
if(token_ids_host(token_id_, i_) == expert_id_)
return i_;
}
throw std::runtime_error("not correct token/expert pair\n");
return -1; // TODO: not correct!!
};
ck_tile::index_t intermediate_size_0 = intermediate_size * (gate_only ? 1 : 2);
ck_tile::index_t intermediate_size_1 = intermediate_size;
ck_tile::HostTensor<AccDataType> out_topk_tokens({tokens, topk, hidden_size});
......@@ -105,11 +94,31 @@ void reference_fused_moe(
if(i_tile >= num_sorted_tiles)
return;
ck_tile::index_t i_expert = sorted_expert_ids_host.mData[i_tile];
ck_tile::index_t i_token = sorted_token_ids_host.mData[i_flatten];
#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
ck_tile::index_t i_token = sorted_token_ids_host.mData[i_flatten];
ck_tile::index_t i_topk = i_token >> 24;
i_token &= 0xffffff;
if(i_token >= tokens)
return;
(void)token_ids_host;
#else
// TODO: better remove this in the future, or modify the token_id value
auto get_topk_id = [&](ck_tile::index_t token_id_, ck_tile::index_t expert_id_) {
for(ck_tile::index_t i_ = 0; i_ < topk; i_++)
{
if(token_ids_host(token_id_, i_) == expert_id_)
return i_;
}
throw std::runtime_error("not correct token/expert pair\n");
return -1; // TODO: not correct!!
};
ck_tile::index_t i_token = sorted_token_ids_host.mData[i_flatten];
if(i_token >= tokens)
return;
ck_tile::index_t i_topk = get_topk_id(i_token, i_expert); // TODO: ugly
auto weight = sorted_weight_host.mData[i_flatten];
#endif
auto weight = sorted_weight_host.mData[i_flatten];
ck_tile::HostTensor<AccDataType> acc_0({1, intermediate_size_0});
// first gemm
......
include/ck_tile/host/reference/reference_gemm.hpp
View file @ 7572a691
......@@ -80,13 +80,14 @@ __global__ void naive_gemm_kernel(ADataType* A,
int b_index = (std::is_same_v<LayoutB, tensor_layout::gemm::ColumnMajor>)
? col * strideB + k
: k * strideB + col;
acc += static_cast<AccDataType>(A[a_index]) * static_cast<AccDataType>(B[b_index]);
acc += ck_tile::type_convert<AccDataType>(A[a_index]) *
ck_tile::type_convert<AccDataType>(B[b_index]);
}
int c_index = (std::is_same_v<LayoutC, tensor_layout::gemm::RowMajor>)
? row * strideC + col
: col * strideC + row;
C[c_index] = acc;
C[c_index] = ck_tile::type_convert<CDataType>(acc);
}
}
......
include/ck_tile/host/reference/reference_moe_sorting.hpp
View file @ 7572a691
......@@ -14,12 +14,15 @@ namespace ck_tile {
template <typename WeightType, typename IndexType = index_t>
CK_TILE_HOST void reference_moe_sorting(const HostTensor<IndexType>& topk_ids,
const HostTensor<WeightType>& weights,
const HostTensor<IndexType>& local_expert_mask,
HostTensor<IndexType>& p_sorted_token_ids,
HostTensor<WeightType>& sorted_weight,
HostTensor<IndexType>& sorted_expert_ids,
index_t& unit_cnt,
const index_t experts,
const index_t unit_size)
const index_t unit_size,
bool local_expert_masking,
bool skip_experts_with_zero_token = true)
{
const index_t num_token = topk_ids.mDesc.get_lengths()[0];
const index_t topk = topk_ids.mDesc.get_lengths()[1];
......@@ -33,8 +36,11 @@ CK_TILE_HOST void reference_moe_sorting(const HostTensor<IndexType>& topk_ids,
#endif
std::vector<std::vector<WeightType>> expert_token_weights(
experts, std::vector<WeightType>(unit_size, 0));
// count number of unit-size slices in this expert
std::vector<IndexType> expert_slices(experts, 1);
// count the tokens used in this expert
std::vector<IndexType> expert_slice_idxs(experts, 0);
// TODO: above 2 buffer seems duplicated
for(index_t t = 0; t < num_token; t++)
{
......@@ -72,8 +78,23 @@ CK_TILE_HOST void reference_moe_sorting(const HostTensor<IndexType>& topk_ids,
IndexType* out_tokens = p_sorted_token_ids.data();
WeightType* out_weights = sorted_weight.data();
IndexType* out_expert_id = sorted_expert_ids.data();
int curr_expert_id = 0;
for(index_t e = 0; e < experts; e++)
{
if(local_expert_masking)
{
if(local_expert_mask(e) == 0)
continue;
}
if(skip_experts_with_zero_token)
{
if(expert_slice_idxs[e] == 0)
{
curr_expert_id++;
continue;
}
}
memcpy(out_tokens, expert_tokens[e].data(), sizeof(index_t) * expert_slices[e] * unit_size);
out_tokens += expert_slices[e] * unit_size;
memcpy(out_weights,
......@@ -83,10 +104,11 @@ CK_TILE_HOST void reference_moe_sorting(const HostTensor<IndexType>& topk_ids,
for(index_t s = 0; s < expert_slices[e]; s++)
{
out_expert_id[s] = e;
out_expert_id[s] = curr_expert_id;
unit_cnt++;
}
out_expert_id += expert_slices[e];
curr_expert_id++;
}
unit_cnt *= unit_size;
return;
......
include/ck_tile/host/reference/reference_rmsnorm2d_fwd.hpp
View file @ 7572a691
......@@ -8,16 +8,40 @@
namespace ck_tile {
// Note: for simplicity, each functor only care about single M
struct reference_rmsnorm2d_default_epilogue
{
template <typename OutDataType, typename AccDataType>
void operator()(int m, HostTensor<OutDataType>& o, const HostTensor<AccDataType>& acc)
{
const int N = acc.mDesc.get_lengths()[1];
for(int n = 0; n < N; ++n)
{
o(m, n) = ck_tile::type_convert<OutDataType>(acc(m, n));
}
}
template <typename OutDataType, typename AccDataType>
auto operator()(int m, const HostTensor<AccDataType>& acc)
{
HostTensor<OutDataType> o(acc.get_lengths(), acc.get_strides());
operator()(m, o, acc);
return o;
}
};
template <typename XDataType,
typename GammaDataType,
typename ComputeDataType,
typename YDataType,
typename InvRmsDataType>
typename InvRmsDataType,
typename Epilogue = reference_rmsnorm2d_default_epilogue>
void reference_rmsnorm2d_fwd(const HostTensor<XDataType>& x_m_n,
const HostTensor<GammaDataType>& gamma_n,
HostTensor<YDataType>& y_m_n,
HostTensor<InvRmsDataType>& invRms_m,
ComputeDataType epsilon)
ComputeDataType epsilon,
Epilogue epilogue_functor = {})
{
auto rmsnorm2d_fwd_func = [&](auto m) {
const int N = x_m_n.mDesc.get_lengths()[1];
......@@ -37,13 +61,15 @@ void reference_rmsnorm2d_fwd(const HostTensor<XDataType>& x_m_n,
if constexpr(!std::is_same_v<InvRmsDataType, ck_tile::null_type>)
invRms_m(m) = ck_tile::type_convert<InvRmsDataType>(divisor);
HostTensor<ComputeDataType> acc(x_m_n.get_lengths(), x_m_n.get_strides());
for(int n = 0; n < N; ++n)
{
ComputeDataType x = ck_tile::type_convert<ComputeDataType>(x_m_n(m, n));
ComputeDataType gamma = ck_tile::type_convert<ComputeDataType>(gamma_n(n));
auto y = x * divisor * gamma;
y_m_n(m, n) = ck_tile::type_convert<YDataType>(y);
acc(m, n) = x * divisor * gamma;
}
epilogue_functor(m, y_m_n, acc);
};
make_ParallelTensorFunctor(rmsnorm2d_fwd_func, invRms_m.mDesc.get_lengths()[0])(
......
include/ck_tile/ops/add_rmsnorm2d_rdquant.hpp
View file @ 7572a691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -10,3 +10,4 @@
#include "ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_three_pass.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/ops/common/utils.hpp"
include/ck_tile/ops/batched_transpose.hpp 0 → 100644
View file @ 7572a691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/ops/batched_transpose/kernel/batched_transpose_kernel.hpp"
#include "ck_tile/ops/batched_transpose/pipeline/batched_transpose_pipeline.hpp"
#include "ck_tile/ops/batched_transpose/pipeline/batched_transpose_policy.hpp"
#include "ck_tile/ops/batched_transpose/pipeline/batched_transpose_problem.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/ops/common/utils.hpp"
include/ck_tile/ops/batched_transpose/kernel/batched_transpose_kernel.hpp 0 → 100644
View file @ 7572a691
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/common.hpp"
#include "ck_tile/ops/elementwise.hpp"
#include "ck_tile/host/hip_check_error.hpp"
#include <string>
#include <type_traits>
namespace ck_tile {
struct BatchedTransposeHostArgs
{
const void* p_input;
void* p_output;
index_t batch;
index_t height;
index_t width;
// index_t dim_blocks;
index_t dim_stride;
index_t dim_block_h;
index_t dim_block_w;
};
template <typename Pipeline_>
struct BatchedTransposeKernel
{
using Pipeline = remove_cvref_t<Pipeline_>;
using Problem = remove_cvref_t<typename Pipeline::Problem>;
using Type = typename Problem::InputType;
struct BatchedTransposeKargs
{
const void* p_input;
void* p_output;
index_t batch;
index_t height;
index_t width;
index_t dim_stride;
};
using Kargs = BatchedTransposeKargs;
using Hargs = BatchedTransposeHostArgs;
CK_TILE_HOST static constexpr auto GridSize(const Hargs& h)
{
size_t grid_size_x = (h.width + h.dim_block_w - 1) / h.dim_block_w;
size_t grid_size_y = (h.height + h.dim_block_h - 1) / h.dim_block_h;
size_t grid_size_z = h.batch;
return dim3(grid_size_x, grid_size_y, grid_size_z);
}
CK_TILE_HOST static constexpr auto MakeKargs(const Hargs& h)
{
Kargs k;
k.p_input = h.p_input;
k.p_output = h.p_output;
k.batch = h.batch;
k.height = h.height;
k.width = h.width;
k.dim_stride = h.dim_stride;
return k;
}
CK_TILE_HOST_DEVICE static constexpr auto BlockSize() { return Problem::kBlockSize; }
CK_TILE_DEVICE void operator()(Kargs kargs) const
{
static constexpr ck_tile::index_t kMPerBlock = Problem::kMPerBlock;
static constexpr ck_tile::index_t kNPerBlock = Problem::kNPerBlock;
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
static constexpr ck_tile::index_t kMPerThread = Problem::kMPerThread;
static constexpr ck_tile::index_t kNPerThread = Problem::kNPerThread;
static_assert(kMPerThread == 1 && kNPerThread == 1);
const auto iDim = blockIdx.z;
const auto x_m_n = [&]() {
const auto x_dram_naive = make_naive_tensor_view<address_space_enum::global>(
static_cast<const Type*>(kargs.p_input) + iDim * kargs.dim_stride,
make_tuple(kargs.height, kargs.width),
make_tuple(kargs.width, 1),
number<kNPerThread>{}, // TODO thread load value
number<1>{});
return pad_tensor_view(x_dram_naive,
make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}),
sequence<kPadM, kPadN>{});
}();
const auto iM = __builtin_amdgcn_readfirstlane(blockIdx.x * kMPerBlock);
const auto iN = __builtin_amdgcn_readfirstlane(blockIdx.y * kNPerBlock);
const auto y_n_m = [&]() {
const auto y_dram_naive = make_naive_tensor_view<address_space_enum::global>(
static_cast<Type*>(kargs.p_output) + iDim * kargs.dim_stride,
make_tuple(kargs.width, kargs.height),
make_tuple(kargs.height, 1),
number<kMPerThread>{},
number<1>{});
return pad_tensor_view(y_dram_naive,
make_tuple(number<kNPerBlock>{}, number<kMPerBlock>{}),
sequence<kPadN, kPadM>{});
}();
auto x_block_window =
make_tile_window(x_m_n,
make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}),
{static_cast<ck_tile::index_t>(iM * kMPerBlock),
static_cast<ck_tile::index_t>(iN * kNPerBlock)});
auto y_block_window =
make_tile_window(y_n_m,
make_tuple(number<kNPerBlock>{}, number<kMPerBlock>{}),
{static_cast<ck_tile::index_t>(iN * kNPerBlock),
static_cast<ck_tile::index_t>(iM * kMPerBlock)});
Pipeline{}(x_block_window, y_block_window);
}
};
} // namespace ck_tile
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