Skip to content

GitLab

Unverified Commit 2a30cfdd authored by arai713's avatar arai713 Committed by GitHub
Browse files

Merge branch 'develop' into codegen-enable-hiprtc

parents 9533a172 78195ccc
Hide whitespace changes
Inline Side-by-side
Showing with 1608 additions and 385 deletions
include/ck_tile/README.md
View file @ 2a30cfdd
# ck_tile
[Back to the main page](../../README.md)
# Composable Kernel Tile
## concept
`ck_tile` provides a programming model with templated abstractions to enable users to implement performance-critical kernels for machine learning workloads. introduces following basic concepts to help users building your own operator
- tensor coordinate transformation, this is the core concept of layout/index transform abstraction in both compiler time and run time.
......@@ -44,5 +45,8 @@ our implementation of different device operators.
**[ops/epilogue]**
epilogue part of our kernel. We may extend this epilogue part to let users to build their own cutomized epilogues.
**[ref]**
reference implementation of cpu or gpu. This folder is supposed to include a specific header on demand.
## examples
currently we put all ck_tile related example under [/example/ck_tile](/example/ck_tile/) folder. Please check each example's subfolder.
include/ck_tile/core.hpp
View file @ 2a30cfdd
// 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
......@@ -7,6 +7,7 @@
#include "ck_tile/core/algorithm/coordinate_transform.hpp"
#include "ck_tile/core/algorithm/indexing_adaptor.hpp"
#include "ck_tile/core/algorithm/space_filling_curve.hpp"
#include "ck_tile/core/algorithm/static_encoding_pattern.hpp"
#include "ck_tile/core/arch/amd_buffer_addressing.hpp"
#include "ck_tile/core/arch/arch.hpp"
#include "ck_tile/core/arch/generic_memory_space_atomic.hpp"
......@@ -31,6 +32,7 @@
#include "ck_tile/core/numeric/math.hpp"
#include "ck_tile/core/numeric/null_type.hpp"
#include "ck_tile/core/numeric/numeric.hpp"
#include "ck_tile/core/numeric/pk_int4.hpp"
#include "ck_tile/core/numeric/type_convert.hpp"
#include "ck_tile/core/numeric/vector_type.hpp"
#include "ck_tile/core/tensor/buffer_view.hpp"
......@@ -53,6 +55,7 @@
#include "ck_tile/core/tensor/tile_window.hpp"
#include "ck_tile/core/tensor/tile_window_linear.hpp"
#include "ck_tile/core/tensor/tile_window_utils.hpp"
#include "ck_tile/core/tensor/transpose_tile.hpp"
#include "ck_tile/core/tensor/update_tile.hpp"
#include "ck_tile/core/utility/bit_cast.hpp"
#include "ck_tile/core/utility/functional.hpp"
......
include/ck_tile/core/algorithm/static_encoding_pattern.hpp 0 → 100644
View file @ 2a30cfdd
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core/arch/arch.hpp"
#include "ck_tile/core/config.hpp"
#include "ck_tile/core/container/sequence.hpp"
#include "ck_tile/core/container/tuple.hpp"
#include "ck_tile/core/numeric/integer.hpp"
#include "ck_tile/core/tensor/tile_distribution.hpp"
#include "ck_tile/core/tensor/tile_distribution_encoding.hpp"
namespace ck_tile {
/**
* @brief Enumeration describing static tile distribution patterns.
*
*/
enum struct tile_distribution_pattern
{
/**
* @brief Thread raked pattern.
*
*/
thread_raked,
/**
* @brief Warp raked pattern.
*
*/
warp_raked,
/**
* @brief Block raked pattern - aka linear.
*
*/
block_raked,
};
struct TileDistributionEncodingPattern
{
};
/**
* @brief Class creating 2D static tile distribution with different load/store patterns.
*
* @note We always assume that Tile is YPerTile x XPerTile where X dim (rightmost)
* is contiguous and we can do vector load on this dimension.
*
* @tparam BlockSize Number of threads in a workgroup.
* @tparam YPerTile The tile size of outer/leftmost dimension.
* @tparam XPerTile The tile size of inner/rightmost dimension (contiguous).
* @tparam VecSize The vector access size.
* @tparam DistributionPattern The enumeration describing used access pattern.
*/
template <index_t BlockSize,
index_t YPerTile,
index_t XPerTile,
index_t VecSize,
tile_distribution_pattern DistributionPattern>
struct TileDistributionEncodingPattern2D : public TileDistributionEncodingPattern
{
};
// Thread raked
template <index_t BlockSize, index_t YPerTile, index_t XPerTile, index_t VecSize>
struct TileDistributionEncodingPattern2D<BlockSize,
YPerTile,
XPerTile,
VecSize,
tile_distribution_pattern::thread_raked>
: public TileDistributionEncodingPattern
{
// TODO: make pattern where below condition does not need to hold - GGemmMultiDSplitk!
static_assert(XPerTile % VecSize == 0, "XPerTile must be a multiple of VecSize!");
static constexpr index_t warp_size = get_warp_size();
static constexpr index_t num_warps = BlockSize / get_warp_size();
static constexpr index_t X1 = VecSize;
static constexpr index_t X0 = XPerTile / X1; // # of threads in X dim
// # of rows in Y dim accessed by single wavefront in one iteration
static constexpr index_t Y1 = warp_size / X0;
static_assert(X0 * Y1 == warp_size, "X0 * Y1 must cover whole wavefront!");
static constexpr index_t Y0 = num_warps;
// YPerWarp = YPerTile / Y0;
// Y2 = YPerWarp / Y1;
static constexpr index_t Y2 = YPerTile / (Y1 * Y0); // # of iters within wavefront
static_assert(X0 * Y1 * Y0 == BlockSize, "X0 * warp_ys * Y0 must cover whole workgroup!");
static_assert(Y0 * Y1 * Y2 == YPerTile, "Y0, Y1, Y2 must cover whole YPerTile");
CK_TILE_HOST_DEVICE static constexpr auto Make2DStaticTileDistribution()
{
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<Y0, Y1, Y2>, sequence<X0, X1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<0>, sequence<1, 0>>,
sequence<1, 2>,
sequence<2, 1>>{});
}
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffled2DStaticTileDistribution()
{
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<X0, X1>, sequence<Y0, Y1, Y2>>,
tuple<sequence<2>, sequence<2, 1>>,
tuple<sequence<0>, sequence<1, 0>>,
sequence<1, 2>,
sequence<1, 2>>{});
}
};
// Warp raked
template <index_t BlockSize, index_t YPerTile, index_t XPerTile, index_t VecSize>
struct TileDistributionEncodingPattern2D<BlockSize,
YPerTile,
XPerTile,
VecSize,
tile_distribution_pattern::warp_raked>
: public TileDistributionEncodingPattern
{
static_assert(XPerTile % VecSize == 0, "XPerTile must be a multiple of VecSize!");
static constexpr index_t warp_size = get_warp_size();
static constexpr index_t num_warps = BlockSize / get_warp_size();
static constexpr index_t X1 = VecSize;
static constexpr index_t X0 = XPerTile / X1; // # of threads in X dim
static constexpr index_t Y2 = warp_size / X0; // # of rows in Y dim to cover whole wavefront
static_assert(X0 * Y2 == warp_size, "X0 * Y2 must cover whole wavefront!");
static constexpr index_t Y0 = num_warps;
static_assert(X0 * Y2 * Y0 == BlockSize, "X0 * Y2 * Y1 must cover whole workgroup!");
static constexpr index_t Y1 = YPerTile / (Y2 * Y0); // # of iters within wavefront
static_assert(Y0 * Y1 * Y2 == YPerTile, "Y0, Y1, Y2 must cover whole YPerTile");
CK_TILE_HOST_DEVICE static constexpr auto Make2DStaticTileDistribution()
{
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<Y0, Y1, Y2>, sequence<X0, X1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<0>, sequence<2, 0>>,
sequence<1, 2>,
sequence<1, 1>>{});
}
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffled2DStaticTileDistribution()
{
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<X0, X1>, sequence<Y0, Y1, Y2>>,
tuple<sequence<2>, sequence<2, 1>>,
tuple<sequence<0>, sequence<2, 0>>,
sequence<1, 2>,
sequence<1, 1>>{});
}
};
// Block raked
template <index_t BlockSize, index_t YPerTile, index_t XPerTile, index_t VecSize>
struct TileDistributionEncodingPattern2D<BlockSize,
YPerTile,
XPerTile,
VecSize,
tile_distribution_pattern::block_raked>
: public TileDistributionEncodingPattern
{
// TODO: make pattern where below condition does not need to hold - GGemmMultiDSplitk!
static_assert(XPerTile % VecSize == 0, "XPerTile must be a multiple of VecSize!");
static constexpr index_t warp_size = get_warp_size();
static constexpr index_t num_warps = BlockSize / get_warp_size();
static constexpr index_t X1 = VecSize;
static constexpr index_t X0 = XPerTile / X1; // # of threads in X dim
static constexpr index_t Y2 = warp_size / X0; // # of rows in Y dim to cover whole wavefront
static_assert(X0 * Y2 == warp_size, "X0 * Y2 must cover whole wavefront!");
static constexpr index_t Y1 = num_warps;
static_assert(X0 * Y2 * Y1 == BlockSize, "X0 * Y2 * Y1 must cover whole workgroup!");
static constexpr index_t Y0 = YPerTile / (Y2 * Y1); // # of iters
static_assert(Y0 * Y1 * Y2 == YPerTile, "Y0, Y1, Y2 must cover whole YPerTile");
CK_TILE_HOST_DEVICE static constexpr auto Make2DStaticTileDistribution()
{
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<Y0, Y1, Y2>, sequence<X0, X1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffled2DStaticTileDistribution()
{
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<X0, X1>, sequence<Y0, Y1, Y2>>,
tuple<sequence<2>, sequence<2, 1>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<1, 0>>{});
}
};
} // namespace ck_tile
include/ck_tile/core/arch/amd_buffer_addressing.hpp
View file @ 2a30cfdd
......@@ -1303,8 +1303,8 @@ CK_TILE_DEVICE thread_buffer<T, N> amd_buffer_load_impl(int32x4_t src_wave_buffe
static_assert(
(std::is_same<T, double>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(std::is_same<T, float>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(std::is_same<T, fp16_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(std::is_same<T, bf16_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(std::is_same<T, fp16_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(std::is_same<T, bf16_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(std::is_same<T, int32_t>::value &&
(N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(std::is_same<T, fp8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
......
include/ck_tile/core/arch/arch.hpp
View file @ 2a30cfdd
// 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
......@@ -12,18 +12,37 @@
namespace ck_tile {
enum struct address_space_enum
template <typename, bool>
struct safe_underlying_type;
template <typename T>
struct safe_underlying_type<T, true>
{
using type = std::underlying_type_t<T>;
};
template <typename T>
struct safe_underlying_type<T, false>
{
using type = void;
};
template <typename T>
using safe_underlying_type_t = typename safe_underlying_type<T, std::is_enum<T>::value>::type;
enum struct address_space_enum : std::uint16_t
{
generic,
generic = 0,
global,
lds,
sgpr,
vgpr,
constant,
vgpr
};
enum struct memory_operation_enum
enum struct memory_operation_enum : std::uint16_t
{
set,
set = 0,
atomic_add,
atomic_max,
add
......@@ -109,4 +128,30 @@ CK_TILE_DEVICE void s_nop(index_t cnt = 0)
#endif
}
#define CK_CONSTANT_ADDRESS_SPACE \
__attribute__((address_space( \
static_cast<safe_underlying_type_t<address_space_enum>>(address_space_enum::constant))))
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/arch/generic_memory_space_atomic.hpp
View file @ 2a30cfdd
// 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/core/numeric/vector_type.hpp"
......@@ -8,16 +8,75 @@
namespace ck_tile {
CK_TILE_HOST_DEVICE bf16_t add_bf16_t(const bf16_t& a, const bf16_t& b)
template <typename T, typename ComputeType>
CK_TILE_HOST_DEVICE T add(const T& a, const T& b)
{
return type_convert<bf16_t>(type_convert<float>(a) + type_convert<float>(b));
return type_convert<T>(type_convert<ComputeType>(a) + type_convert<ComputeType>(b));
}
CK_TILE_HOST_DEVICE bf16x2_t add_bf16x2_t(const bf16x2_t& a, const bf16x2_t& b)
{
bf16x2_t rtn;
rtn[0] = add_bf16_t(a[0], b[0]);
rtn[1] = add_bf16_t(a[1], b[1]);
rtn[0] = add<bf16_t, float>(a[0], b[0]);
rtn[1] = add<bf16_t, float>(a[1], b[1]);
return rtn;
}
CK_TILE_HOST_DEVICE bf16x4_t add_bf16x4_t(const bf16x4_t& a, const bf16x4_t& b)
{
bf16x4_t rtn;
rtn[0] = add<bf16_t, float>(a[0], b[0]);
rtn[1] = add<bf16_t, float>(a[1], b[1]);
rtn[2] = add<bf16_t, float>(a[2], b[2]);
rtn[3] = add<bf16_t, float>(a[3], b[3]);
return rtn;
}
CK_TILE_HOST_DEVICE fp8x4_t add_fp8x4_t(const fp8x4_t& a, const fp8x4_t& b)
{
fp8x4_t rtn;
rtn[0] = add<fp8_t, float>(a[0], b[0]);
rtn[1] = add<fp8_t, float>(a[1], b[1]);
rtn[2] = add<fp8_t, float>(a[2], b[2]);
rtn[3] = add<fp8_t, float>(a[3], b[3]);
return rtn;
}
CK_TILE_HOST_DEVICE fp8x8_t add_fp8x8_t(const fp8x8_t& a, const fp8x8_t& b)
{
fp8x8_t rtn;
rtn[0] = add<fp8_t, float>(a[0], b[0]);
rtn[1] = add<fp8_t, float>(a[1], b[1]);
rtn[2] = add<fp8_t, float>(a[2], b[2]);
rtn[3] = add<fp8_t, float>(a[3], b[3]);
rtn[4] = add<fp8_t, float>(a[4], b[4]);
rtn[5] = add<fp8_t, float>(a[5], b[5]);
rtn[6] = add<fp8_t, float>(a[6], b[6]);
rtn[7] = add<fp8_t, float>(a[7], b[7]);
return rtn;
}
CK_TILE_HOST_DEVICE bf8x4_t add_bf8x4_t(const bf8x4_t& a, const bf8x4_t& b)
{
bf8x4_t rtn;
rtn[0] = add<bf8_t, float>(a[0], b[0]);
rtn[1] = add<bf8_t, float>(a[1], b[1]);
rtn[2] = add<bf8_t, float>(a[2], b[2]);
rtn[3] = add<bf8_t, float>(a[3], b[3]);
return rtn;
}
CK_TILE_HOST_DEVICE bf8x8_t add_bf8x8_t(const bf8x8_t& a, const bf8x8_t& b)
{
bf8x8_t rtn;
rtn[0] = add<bf8_t, float>(a[0], b[0]);
rtn[1] = add<bf8_t, float>(a[1], b[1]);
rtn[2] = add<bf8_t, float>(a[2], b[2]);
rtn[3] = add<bf8_t, float>(a[3], b[3]);
rtn[4] = add<bf8_t, float>(a[4], b[4]);
rtn[5] = add<bf8_t, float>(a[5], b[5]);
rtn[6] = add<bf8_t, float>(a[6], b[6]);
rtn[7] = add<bf8_t, float>(a[7], b[7]);
return rtn;
}
......@@ -59,6 +118,192 @@ CK_TILE_DEVICE void atomic_add<bf16x2_t>(bf16x2_t* p_dst, const bf16x2_t& x)
} while(cur_v.u32 != old_v);
}
template <>
CK_TILE_DEVICE void atomic_add<bf16x4_t>(bf16x4_t* p_dst, bf16x4_t const& x)
{
// Union to treat the pointer as either bf16x4_t* or uint64_t*:
union U64BF164_ADDR
{
uint64_t* u64_a;
bf16x4_t* bf164_a;
};
// Union to treat the data as either bf16x4_t or 64-bit integer
union U64BF164
{
uint64_t u64;
bf16x4_t bf164;
};
U64BF164_ADDR addr;
addr.bf164_a = p_dst; // interpret p_dst as a 64-bit location
// First read (non-atomic) of the old value
U64BF164 cur_v;
cur_v.u64 = *addr.u64_a;
U64BF164 new_v_union;
uint64_t old_v, new_v;
do
{
// old 64 bits
old_v = cur_v.u64;
// Add elementwise in bf16
new_v_union.bf164 = add_bf16x4_t(cur_v.bf164, x);
new_v = new_v_union.u64;
// Attempt the 64-bit CAS
cur_v.u64 = atomicCAS(addr.u64_a, old_v, new_v);
} while(cur_v.u64 != old_v);
}
template <>
CK_TILE_DEVICE void atomic_add<fp8x4_t>(fp8x4_t* p_dst, const fp8x4_t& x)
{
union U32FP84_ADDR
{
uint32_t* u32_a;
fp8x4_t* fp84_a;
};
union U32FP84
{
uint32_t u32;
fp8x4_t fp84;
};
U32FP84_ADDR dword_addr;
U32FP84 cur_v;
U32FP84 new_;
uint32_t old_v, new_v;
dword_addr.fp84_a = p_dst;
cur_v.u32 = *dword_addr.u32_a;
do
{
old_v = cur_v.u32;
new_.fp84 = add_fp8x4_t(cur_v.fp84, x);
new_v = new_.u32;
cur_v.u32 = atomicCAS(dword_addr.u32_a, old_v, new_v);
} while(cur_v.u32 != old_v);
}
template <>
CK_TILE_DEVICE void atomic_add<bf8x4_t>(bf8x4_t* p_dst, const bf8x4_t& x)
{
union U32BF84_ADDR
{
uint32_t* u32_a;
bf8x4_t* bf84_a;
};
union U32BF84
{
uint32_t u32;
bf8x4_t bf84;
};
U32BF84_ADDR dword_addr;
U32BF84 cur_v;
U32BF84 new_;
uint32_t old_v, new_v;
dword_addr.bf84_a = p_dst;
cur_v.u32 = *dword_addr.u32_a;
do
{
old_v = cur_v.u32;
new_.bf84 = add_bf8x4_t(cur_v.bf84, x);
new_v = new_.u32;
cur_v.u32 = atomicCAS(dword_addr.u32_a, old_v, new_v);
} while(cur_v.u32 != old_v);
}
//
// Atomic add for fp8x8_t
//
template <>
CK_TILE_DEVICE void atomic_add<fp8x8_t>(fp8x8_t* p_dst, fp8x8_t const& x)
{
// Union for addressing 64 bits as either "fp8x8_t" or a 64-bit integer.
union U64FP88_ADDR
{
uint64_t* u64_a; // pointer to 64-bit integer
fp8x8_t* fp88_a; // pointer to fp8x8_t
};
union U64FP88
{
uint64_t u64;
fp8x8_t fp88;
};
U64FP88_ADDR dword_addr;
U64FP88 cur_v;
U64FP88 new_v_union;
uint64_t old_v, new_v;
// Point to the destination as both fp8x8_t* and uint64_t*.
dword_addr.fp88_a = p_dst;
// Initial read of 64 bits from memory
cur_v.u64 = *dword_addr.u64_a;
do
{
old_v = cur_v.u64;
// Add each fp8 element using your add_fp8x8_t(...) routine
new_v_union.fp88 = add_fp8x8_t(cur_v.fp88, x);
new_v = new_v_union.u64;
// Attempt 64-bit CAS
cur_v.u64 = atomicCAS(dword_addr.u64_a, old_v, new_v);
} while(cur_v.u64 != old_v);
}
//
// Atomic add for bf8x8_t
//
template <>
CK_TILE_DEVICE void atomic_add<bf8x8_t>(bf8x8_t* p_dst, bf8x8_t const& x)
{
union U64BF88_ADDR
{
uint64_t* u64_a;
bf8x8_t* bf88_a;
};
union U64BF88
{
uint64_t u64;
bf8x8_t bf88;
};
U64BF88_ADDR dword_addr;
U64BF88 cur_v;
U64BF88 new_v_union;
uint64_t old_v, new_v;
dword_addr.bf88_a = p_dst;
// Read the original 64 bits
cur_v.u64 = *dword_addr.u64_a;
do
{
old_v = cur_v.u64;
// Add each bf8 element using your add_bf8x8_t(...) routine
new_v_union.bf88 = add_bf8x8_t(cur_v.bf88, x);
new_v = new_v_union.u64;
// 64-bit CAS loop
cur_v.u64 = atomicCAS(dword_addr.u64_a, old_v, new_v);
} while(cur_v.u64 != old_v);
}
template <typename T, index_t N>
CK_TILE_DEVICE void atomic_add_g(T* p_dst, const thread_buffer<T, N>& x)
{
......@@ -66,8 +311,10 @@ CK_TILE_DEVICE void atomic_add_g(T* p_dst, const thread_buffer<T, N>& x)
(std::is_same<T, uint32_t>::value && (N == 1)) ||
(std::is_same<T, float>::value && (N == 1 || N == 2)) ||
(std::is_same<T, double>::value && (N == 1 || N == 2)) ||
(std::is_same<T, bf16_t>::value && (N == 2 || N == 4)),
"wrong! not implemented");
(std::is_same<T, bf16_t>::value && (N == 2 || N == 4 || N == 8)) ||
(std::is_same<T, fp8_t>::value && (N == 4 || N == 8 || N == 16)) ||
(std::is_same<T, bf8_t>::value && (N == 4 || N == 8 || N == 16)),
"The granularity of the thread buffer is unsupported on the hardware!");
constexpr auto I0 = number<0>{};
constexpr auto I1 = number<1>{};
......@@ -118,9 +365,45 @@ CK_TILE_DEVICE void atomic_add_g(T* p_dst, const thread_buffer<T, N>& x)
}
else if constexpr(N == 4)
{
atomic_add(c_style_pointer_cast<bf16x2_t*>(p_dst), x.template get_as<bf16x2_t>()[I0]);
atomic_add(c_style_pointer_cast<bf16x2_t*>(p_dst) + 1,
x.template get_as<bf16x2_t>()[I1]);
atomic_add(c_style_pointer_cast<bf16x4_t*>(p_dst), x.template get_as<bf16x4_t>()[I0]);
}
else if constexpr(N == 8)
{
atomic_add(c_style_pointer_cast<bf16x4_t*>(p_dst), x.template get_as<bf16x4_t>()[I0]);
atomic_add(c_style_pointer_cast<bf16x4_t*>(p_dst) + 1,
x.template get_as<bf16x4_t>()[I1]);
}
}
else if constexpr(std::is_same<T, fp8_t>::value)
{
if constexpr(N == 4)
{
atomic_add(c_style_pointer_cast<fp8x4_t*>(p_dst), x.template get_as<fp8x4_t>()[I0]);
}
if constexpr(N == 8)
{
atomic_add(c_style_pointer_cast<fp8x8_t*>(p_dst), x.template get_as<fp8x8_t>()[I0]);
}
if constexpr(N == 16)
{
atomic_add(c_style_pointer_cast<fp8x8_t*>(p_dst), x.template get_as<fp8x8_t>()[I0]);
atomic_add(c_style_pointer_cast<fp8x8_t*>(p_dst) + 1, x.template get_as<fp8x8_t>()[I1]);
}
}
else if constexpr(std::is_same<T, bf8_t>::value)
{
if constexpr(N == 4)
{
atomic_add(c_style_pointer_cast<bf8x4_t*>(p_dst), x.template get_as<bf8x4_t>()[I0]);
}
if constexpr(N == 8)
{
atomic_add(c_style_pointer_cast<bf8x8_t*>(p_dst), x.template get_as<bf8x8_t>()[I0]);
}
if constexpr(N == 16)
{
atomic_add(c_style_pointer_cast<bf8x8_t*>(p_dst), x.template get_as<bf8x8_t>()[I0]);
atomic_add(c_style_pointer_cast<bf8x8_t*>(p_dst) + 1, x.template get_as<bf8x8_t>()[I1]);
}
}
}
......
include/ck_tile/core/config.hpp
View file @ 2a30cfdd
// 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
#if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
defined(__gfx942__)
defined(__gfx942__) || defined(__gfx950__)
#define __gfx9__
#endif
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
#define __gfx94__
#endif
#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || \
......@@ -144,6 +144,10 @@
#define CK_TILE_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER 1
#endif
#ifndef CK_TILE_USE_PK4_LAYOUT_SHUFFLE
#define CK_TILE_USE_PK4_LAYOUT_SHUFFLE 1
#endif
// buffer atomic add: floating point
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_TILE_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
......@@ -230,3 +234,15 @@
#ifndef CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
#define CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID 1
#endif
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#ifdef CK_TILE_USE_OCP_FP8
#define CK_TILE_USE_OCP_FP8 1
#else
#define CK_TILE_USE_OCP_FP8 0
#endif
#elif defined(__gfx950__) || defined(__gfx12__) // for GPU code
#define CK_TILE_USE_OCP_FP8 1
#else // for GPU code
#define CK_TILE_USE_OCP_FP8 0
#endif
include/ck_tile/core/container/meta_data_buffer.hpp
View file @ 2a30cfdd
......@@ -30,7 +30,7 @@ struct meta_data_buffer
{
constexpr index_t size = sizeof(T);
auto tmp = bit_cast<array<std::byte, size>>(data);
auto tmp = ck_tile::bit_cast<array<std::byte, size>>(data);
for(int i = 0; i < size; i++)
{
......@@ -66,7 +66,7 @@ struct meta_data_buffer
pos++;
}
data = bit_cast<T>(tmp);
data = ck_tile::bit_cast<T>(tmp);
}
return data;
......@@ -86,7 +86,7 @@ struct meta_data_buffer
pos++;
}
auto data = bit_cast<T>(tmp);
auto data = ck_tile::bit_cast<T>(tmp);
return data;
}
......
include/ck_tile/core/container/tuple.hpp
View file @ 2a30cfdd
......@@ -546,7 +546,7 @@ CK_TILE_HOST_DEVICE constexpr auto tuple_reverse(const tuple<Ts...>& t)
using Idx = number<tuple<Ts...>::size() - i - 1>;
return t.at(Idx{});
},
number<tuple<Ts...>::size()()>{});
number<tuple<Ts...>::size()>{});
}
// Reduce tuple values in specific range using Function
......
include/ck_tile/core/numeric/bfloat16.hpp
View file @ 2a30cfdd
// 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.
#include "ck_tile/core/config.hpp"
#include "ck_tile/core/utility/bit_cast.hpp"
......@@ -376,6 +376,16 @@ struct numeric<bfloat16_t>
}
};
template <typename T>
struct numeric_traits;
template <>
struct numeric_traits<bfloat16_t>
{
static constexpr int exp = 8;
static constexpr int mant = 7;
};
#if CK_TILE_USE_CUSTOM_DATA_TYPE
CK_TILE_ARITHMETIC_USING_FLOAT(CK_TILE_HOST_DEVICE, bfloat16_t)
#endif
......
include/ck_tile/core/numeric/float8.hpp
View file @ 2a30cfdd
// 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.
#include "ck_tile/core/config.hpp"
#include "ck_tile/core/utility/bit_cast.hpp"
......@@ -14,6 +14,12 @@
#pragma once
#if(defined(__gfx94__) || defined(__gfx12__)) && __HIP_DEVICE_COMPILE__
#define CK_TILE_FP8_CVT_DEVICE 1
#else
#define CK_TILE_FP8_CVT_DEVICE 0
#endif
namespace ck_tile {
// fp8 rounding modes
......@@ -25,15 +31,26 @@ enum class fp8_rounding_mode
stochastic
};
/**
* \brief FP8 interpretation used in conversion algorithms
*/
enum class fp8_interpretation
{
E4M3_OCP = 0, // OCP FP8 E4M3
E5M2_OCP = 1, // OCP BF8 E5M2
E4M3_FNUZ = 2, // FNUZ FP8 E4M3
E5M2_FNUZ = 3, // FNUZ BF8 E5M2
};
/*
* ______________NANOO_________________ | ______________IEEE________________
* ______________FNUZ_________________ | ______________OCP________________
* e4m3 e5m2 | e4m3 e5m2
* bias : 8 16 | 7 15
* inf : 1.0000.000 1.00000.00 | N/A s.11111.00
* Nan : 1.0000.000 1.00000.00 | s.1111.111 s.11111.{01, 10, 11}
* zero : 0.0000.000 0.00000.00 | s.0000.000 s.00000.00
* Max(norm) : s.1111.111 (240) s.11111.11(57344) | s.1111.110(448) s.11110.11(57344)
* Max(snorm): s.0000.111 s.00000.11 | s.0000.111(448) s.00000.11(57344)
* Max(snorm): s.0000.111 s.00000.11 | s.0000.111 s.00000.11
* 0.0068359375 2.288818e-05 | 0.013671875 4.57763671875e-05
* Min(norm) : s.0001.000 s.00001.00 | s.0001.000 s.00001.00
* 2^-7(0.00078125) 2^-15(3.05176e-05) | 2^-6(0.015625) 2^-14(6.10352e-05)
......@@ -55,10 +72,10 @@ struct alignas(1) float8_e4m3_t
{
static constexpr int exponent = 4;
static constexpr int mantissa = 3;
#if defined(__gfx94__)
static constexpr int bias = 1 << (exponent - 1); // NANOO
#if CK_TILE_USE_OCP_FP8
static constexpr int bias = 7; // OCP
#else
static constexpr int bias = (1 << (exponent - 1)) - 1; // IEEE
static constexpr int bias = 8; // FNUZ
#endif
using raw_type = uint8_t;
raw_type data;
......@@ -113,10 +130,10 @@ struct alignas(1) float8_e5m2_t
{
static constexpr int exponent = 5;
static constexpr int mantissa = 2;
#if defined(__gfx94__)
static constexpr int bias = 1 << (exponent - 1); // NANOO
#if CK_TILE_USE_OCP_FP8
static constexpr int bias = 15; // OCP
#else
static constexpr int bias = (1 << (exponent - 1)) - 1; // IEEE
static constexpr int bias = 16; // FNUZ
#endif
using raw_type = uint8_t;
raw_type data;
......@@ -183,501 +200,727 @@ struct native_t<bf8_t>
};
#else
using fp8_t = _BitInt(8);
using fp8_raw_t = uint8_t;
using bf8_t = unsigned _BitInt(8);
using bf8_raw_t = uint8_t;
#endif
// below is sw fp8 conversion, not utilizing hw instruction
namespace impl {
template <typename T>
struct numeric_traits;
template <typename X, typename Y, bool negative_zero_nan, bool clip, bool stoch>
CK_TILE_HOST_DEVICE Y run_cast_to_f8(X x, uint32_t rng)
template <>
struct numeric_traits<fp8_t>
{
// fp8/bf8 exponent/mantissa layout
constexpr int out_exp = numeric_traits<Y>::exp;
constexpr int out_mant = numeric_traits<Y>::mant;
using bitwise_type = fp8_raw_t;
static constexpr int exp = 4;
static constexpr int mant = 3;
#if CK_TILE_USE_OCP_FP8
static constexpr int bias = 7;
static constexpr fp8_interpretation f8_interpret = fp8_interpretation::E4M3_OCP;
#else
static constexpr int bias = 8;
static constexpr fp8_interpretation f8_interpret = fp8_interpretation::E4M3_FNUZ;
#endif
static constexpr uint8_t abs_mask = 0x7F;
};
// original type exponent/mantissa layout
constexpr int in_exp = numeric_traits<X>::exp;
constexpr int in_mant = numeric_traits<X>::mant;
template <>
struct numeric_traits<bf8_t>
{
using bitwise_type = bf8_raw_t;
int exponent, bias;
uint32_t head, mantissa, sign;
// nan code is same for float and half
#if CK_TILE_USE_CUSTOM_DATA_TYPE
constexpr Y nan_code =
numeric<Y>::quiet_NaN(); // __builtin_bit_cast(Y, static_cast<uint8_t>(0x80));
static constexpr int exp = 5;
static constexpr int mant = 2;
#if CK_TILE_USE_OCP_FP8
static constexpr int bias = 15;
static constexpr fp8_interpretation f8_interpret = fp8_interpretation::E5M2_OCP;
#else
constexpr Y nan_code = 0x80;
static constexpr int bias = 16;
static constexpr fp8_interpretation f8_interpret = fp8_interpretation::E5M2_FNUZ;
#endif
static constexpr uint8_t abs_mask = 0x7F;
};
// below is sw fp8 conversion, not utilizing hw instruction
namespace impl {
template <typename SrcT, typename DstT, bool clip = true, bool stoch = false>
CK_TILE_HOST_DEVICE DstT run_cast_to_f8(SrcT src, unsigned int rng = 0)
{
static_assert(std::is_same<DstT, fp8_t>::value || std::is_same<DstT, bf8_t>::value,
"DstT type must be fp8 or bf8.");
constexpr uint32_t nan_mask = numeric_traits<X>::nan_mask;
constexpr bool is_half = std::is_same<SrcT, half_t>::value;
constexpr bool is_float = std::is_same<SrcT, float>::value;
static_assert(is_half || is_float, "Only half and float can be cast to f8");
// convert to bitwise
using T_bitwise = typename numeric_traits<X>::bitwise_type;
T_bitwise x_bitwise = *(reinterpret_cast<T_bitwise*>(&x));
// fp8/bf8 type exponent/mantissa layout
constexpr int DstT_exp = numeric_traits<DstT>::exp; // exponent width of the destination type
constexpr int DstT_mant = numeric_traits<DstT>::mant; // mantissa width of the destination type
constexpr bool is_fnuz =
(numeric_traits<DstT>::f8_interpret == fp8_interpretation::E4M3_FNUZ) ||
(numeric_traits<DstT>::f8_interpret == fp8_interpretation::E5M2_FNUZ);
// unpack the input, depends on datatype
head = x_bitwise & numeric_traits<X>::head_mask;
mantissa = x_bitwise & numeric_traits<X>::mant_mask;
exponent = (head >> in_mant) & numeric_traits<X>::exp_mask;
sign = head >> (in_exp + in_mant);
bias = numeric_traits<X>::bias;
constexpr int SrcT_exp = numeric_traits<SrcT>::exp;
constexpr int SrcT_mant = numeric_traits<SrcT>::mant;
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);
using SrcT_bitwise = typename numeric_traits<SrcT>::bitwise_type;
SrcT_bitwise src_bitwise = bit_cast<SrcT_bitwise>(src);
if constexpr(negative_zero_nan)
unsigned long long head, mantissa;
int exponent, bias;
unsigned int sign;
unsigned long long fInf, abs_mask;
head = src_bitwise & numeric_traits<SrcT>::head_mask;
mantissa = src_bitwise & numeric_traits<SrcT>::mant_mask;
exponent = (head >> SrcT_mant) & numeric_traits<SrcT>::exp_mask;
sign = head >> (SrcT_exp + SrcT_mant);
bias = numeric_traits<SrcT>::bias;
fInf = numeric_traits<SrcT>::Inf;
abs_mask = numeric_traits<SrcT>::abs_mask;
unsigned int signed_inf = 0;
unsigned int nan = 0;
if constexpr(is_fnuz)
{
if((x_bitwise & nan_mask) == nan_mask)
return nan_code;
signed_inf = clip ? ((sign << 7) + 0x7f) : 0x80;
nan = 0x80;
}
else
{
if((x_bitwise & nan_mask) == nan_mask)
return signed_inf + (mantissa != 0 ? 1 : 0);
if constexpr(DstT_exp == 4)
{ // e4m3
signed_inf = (sign << 7) + (clip ? 0x7e : 0x7f);
}
else
{ // e5m2
signed_inf = (sign << 7) + (clip ? 0x7b : 0x7c);
}
nan = (sign << 7) + 0x7f;
}
// Max values
unsigned long long ifmax = 0;
if constexpr(is_float)
{
if constexpr(DstT_exp == 5)
{
ifmax = 0x47600000;
}
else
{
if constexpr(is_fnuz)
{
ifmax = 0x43700000;
}
else
{
ifmax = 0x43E00000;
}
}
}
else if constexpr(is_half)
{
if constexpr(DstT_exp == 5)
{
ifmax = 0x7B00;
}
else
{
if constexpr(is_fnuz)
{
ifmax = 0x5B80;
}
else
{
ifmax = 0x5F00;
}
}
}
// check if x is 0.0
if(x_bitwise == 0)
return __builtin_bit_cast(Y, static_cast<uint8_t>(0));
// Deal with inf and NaNs
if((src_bitwise & fInf) == fInf)
{
if constexpr(is_fnuz)
return signed_inf;
return mantissa != 0 ? nan : signed_inf;
}
if((src_bitwise & abs_mask) > ifmax)
{
return signed_inf;
}
if(src_bitwise == 0)
{
return 0;
}
// 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
// First need to check if it is normal or denorm as there is a difference of
// implicit 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 again
// 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
// For IEEE bias mode, the bias is 2^(k-1) -1 where k is the width of exponent
// bits
const int f8_bias = (1 << (DstT_exp - 1)) - 1 + (is_fnuz ? 1 : 0);
const int f8_denormal_act_exponent = 1 - f8_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
// f8_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;
int act_exponent, f8_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 */
/* 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 -
exponent_diff = f8_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)
if(act_exponent <= f8_denormal_act_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;
/* 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 implicit 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 = f8_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
{ // 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 += (1ull << SrcT_mant); // Add the implicit 1 into mantissa
}
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. */
bool midpoint = (mantissa & ((1ull << (SrcT_mant - DstT_mant + exponent_diff)) - 1)) ==
(1ull << (SrcT_mant - DstT_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);
bool implicit_one = mantissa & (1ull << SrcT_mant);
// if there is no implicit 1, it means the f8 is denormal and need to adjust
// to denorm exponent
f8_exponent =
(act_exponent + exponent_diff) /*actual f8 exponent*/ + f8_bias - (implicit_one ? 0 : 1);
// Now we have the exponent and mantissa adjusted
unsigned long long drop_mask = (1ull << (SrcT_mant - DstT_mant)) - 1;
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;
mantissa & (1ull << (SrcT_mant -
DstT_mant)); // if the least significant bit that is not truncated is 1
mantissa +=
(stoch ? rng : (midpoint ? (odd ? mantissa : mantissa - 1ull) : mantissa)) & drop_mask;
// Now we deal with overflow
if(out_exponent == 0)
if(f8_exponent == 0)
{
if((1 << in_mant) & mantissa)
if((1ull << SrcT_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
f8_exponent = 1; // denormal overflow to become normal, promote exponent
}
}
else
{
if((1 << (in_mant + 1)) & mantissa)
if((1ull << (SrcT_mant + 1)) & mantissa)
{
mantissa >>= 1;
out_exponent++;
// No need to make 1 implicit now as it will be addressed later
f8_exponent++;
}
}
mantissa >>= (in_mant - out_mant);
mantissa >>= (SrcT_mant - DstT_mant);
if(out_exponent > max_exp)
// above range: quantize to maximum possible float of the same sign
const int max_exp = (1 << DstT_exp) - 1;
if(f8_exponent > max_exp)
{
if(clip)
if constexpr(clip)
{
mantissa = (1 << out_mant) - 1;
out_exponent = max_exp;
mantissa = (1 << DstT_mant) - 1;
f8_exponent = max_exp;
}
else
{
return __builtin_bit_cast(Y, static_cast<uint8_t>(signed_inf));
return signed_inf;
}
}
// check if x is 0.0 or -0.0
if(out_exponent == 0 && mantissa == 0)
return __builtin_bit_cast(
Y, static_cast<uint8_t>(negative_zero_nan ? 0 : (sign << (out_exp + out_mant))));
mantissa &= (1 << out_mant) - 1;
return __builtin_bit_cast(Y,
static_cast<uint8_t>((sign << (out_exp + out_mant)) |
(out_exponent << out_mant) | mantissa));
if(f8_exponent == 0 && mantissa == 0)
return is_fnuz ? 0 : (sign << 7);
mantissa &= (1 << DstT_mant) - 1;
return (sign << 7) | (f8_exponent << DstT_mant) | mantissa;
}
template <typename X, typename Y, bool negative_zero_nan>
CK_TILE_HOST_DEVICE Y run_cast_from_f8(X x)
template <typename SrcT, typename DstT, bool clip = true>
CK_TILE_HOST_DEVICE DstT run_cast_from_f8(SrcT x)
{
// fp8/bf8 exponent/mantissa layout
constexpr int in_exp = numeric_traits<X>::exp;
constexpr int in_mant = numeric_traits<X>::mant;
// resulting type exponent/mantissa layout
constexpr int out_exp = numeric_traits<Y>::exp;
constexpr int out_mant = numeric_traits<Y>::mant;
uint8_t x_raw = __builtin_bit_cast(uint8_t, x);
// prepare the codes
constexpr uint8_t nan_code = 0x80;
Y Inf, NegInf, NaN, Neg0;
using T_bitwise = typename numeric_traits<Y>::bitwise_type;
constexpr T_bitwise Inf_bitwise = numeric_traits<Y>::Inf;
constexpr T_bitwise NegInf_bitwise = numeric_traits<Y>::NegInf;
constexpr T_bitwise NaN_bitwise = numeric_traits<Y>::NaN;
constexpr T_bitwise Neg0_bitwise = numeric_traits<Y>::Neg0;
Inf = *(reinterpret_cast<const Y*>(&Inf_bitwise));
NegInf = *(reinterpret_cast<const Y*>(&NegInf_bitwise));
NaN = *(reinterpret_cast<const Y*>(&NaN_bitwise));
Neg0 = *(reinterpret_cast<const Y*>(&Neg0_bitwise));
// check if x is 0.0
if(x_raw == 0)
return static_cast<Y>(0);
// unpack the input
uint32_t sign = x_raw >> (in_exp + in_mant);
uint32_t mantissa = x_raw & ((1 << in_mant) - 1);
int exponent = (x_raw & 0x7F) >> in_mant;
static_assert(std::is_same<SrcT, fp8_t>::value || std::is_same<SrcT, bf8_t>::value,
"SrcT type must be fp8 or bf8.");
constexpr int SrcT_exp = numeric_traits<SrcT>::exp;
constexpr int SrcT_mant = numeric_traits<SrcT>::mant;
constexpr bool is_fnuz =
(numeric_traits<SrcT>::f8_interpret == fp8_interpretation::E4M3_FNUZ) ||
(numeric_traits<SrcT>::f8_interpret == fp8_interpretation::E5M2_FNUZ);
constexpr bool is_half = std::is_same<DstT, half_t>::value;
constexpr bool is_float = std::is_same<DstT, float>::value;
static_assert(is_half || is_float, "DstT type must be half_t or float.");
// destination type exponent/mantissa layout
constexpr int DstT_exp = numeric_traits<DstT>::exp; // exponent width of the destination type
constexpr int DstT_mant = numeric_traits<DstT>::mant; // mantissa width of the destination type
constexpr DstT fInf = bit_cast<DstT>(numeric_traits<DstT>::Inf);
constexpr DstT fNegInf = bit_cast<DstT>(numeric_traits<DstT>::NegInf);
constexpr DstT fNaN = bit_cast<DstT>(numeric_traits<DstT>::NaN);
constexpr DstT fNeg0 = bit_cast<DstT>(numeric_traits<DstT>::Neg0);
DstT fmax{0}, fmin{0};
// Max number in e5m2 57344
if constexpr(is_half)
{
fmax = bit_cast<DstT>(static_cast<typename numeric_traits<DstT>::bitwise_type>(0x7B00));
fmin = bit_cast<DstT>(static_cast<typename numeric_traits<DstT>::bitwise_type>(0xFB00));
}
else if constexpr(is_float)
{
fmax = bit_cast<DstT>(static_cast<typename numeric_traits<DstT>::bitwise_type>(0x47600000));
fmin = bit_cast<DstT>(static_cast<typename numeric_traits<DstT>::bitwise_type>(0xC7600000));
}
constexpr int exp_low_cutoff =
(1 << (out_exp - 1)) - (1 << (in_exp - 1)) + 1 - (negative_zero_nan ? 1 : 0);
T_bitwise retval;
if(x == 0)
{
return 0;
}
if constexpr(negative_zero_nan)
unsigned long long sign = x >> 7;
unsigned long long mantissa = x & ((1 << SrcT_mant) - 1);
int exponent = (x & 0x7F) >> SrcT_mant;
if constexpr(is_fnuz)
{
if(x_raw == nan_code)
return NaN;
if(x == 0x80)
{
return fNaN;
}
}
else
{
if(x_raw == nan_code)
return Neg0;
if(exponent == ((1 << in_exp) - 1))
return (mantissa == 0) ? (sign ? NegInf : Inf) : NaN;
if(x == 0x80)
{
return fNeg0;
}
if constexpr(SrcT_exp == 4)
{ // e4m3
if((x & 0x7F) == 0x7F)
{
return fNaN;
}
}
else if((x & 0x7C) == 0x7C)
{ // e5m2
if((x & 0x3) == 0)
{
if constexpr(clip)
{
return sign ? fmin : fmax;
}
return sign ? fNegInf : fInf;
}
return fNaN;
}
}
if((numeric_traits<Y>::mant == 10) && (numeric_traits<X>::mant == 2) && !negative_zero_nan)
typename numeric_traits<DstT>::bitwise_type retval;
if constexpr(SrcT_exp == 5 && is_half && !is_fnuz)
{
retval = x_raw;
retval <<= 8;
return *(reinterpret_cast<const Y*>(&retval));
retval = x << 8;
return bit_cast<DstT>(retval);
}
const int exp_low_cutoff =
(1 << (DstT_exp - 1)) - (1 << (SrcT_exp - 1)) + 1 - (is_fnuz ? 1 : 0);
// subnormal input
if(exponent == 0)
{
// guaranteed mantissa!=0 since cases 0x0 and 0x80 are handled above
int sh = 1 + clz(mantissa) - (32 - in_mant);
int sh = 1 + clz(mantissa) - (32 - SrcT_mant);
mantissa <<= sh;
exponent += 1 - sh;
mantissa &= ((1 << in_mant) - 1);
mantissa &= ((1ull << SrcT_mant) - 1);
}
exponent += exp_low_cutoff - 1;
mantissa <<= out_mant - in_mant;
mantissa <<= DstT_mant - SrcT_mant;
// subnormal output (occurs when T=half, we=5, negative_zero_nan=true)
// subnormal output (occurs when DstT is half_t, we=5, is_fnuz=true)
if(exponent <= 0)
{
mantissa |= 1 << out_mant;
mantissa |= 1 << DstT_mant;
mantissa >>= 1 - exponent;
exponent = 0;
}
retval = (sign << (out_exp + out_mant)) | (exponent << out_mant) | mantissa;
return *(reinterpret_cast<const Y*>(&retval));
}
template <typename X, typename Y, bool negative_zero_nan, bool clip, bool stoch>
CK_TILE_HOST_DEVICE Y cast_to_f8(X x, uint32_t rng)
{
// check datatypes
constexpr bool is_half = std::is_same<X, half_t>::value;
constexpr bool is_float = std::is_same<X, float>::value;
static_assert(is_half || is_float, "Only half and float can be casted.");
retval = (sign << (DstT_exp + DstT_mant)) | (exponent << DstT_mant) | mantissa;
return run_cast_to_f8<X, Y, negative_zero_nan, clip, stoch>(x, rng);
return bit_cast<DstT>(retval);
}
template <typename X, typename Y, bool negative_zero_nan>
CK_TILE_HOST_DEVICE Y cast_from_f8(X x)
template <typename X, typename Y, bool clip, bool stoch>
CK_TILE_HOST_DEVICE Y cast_to_f8(X x, uint32_t rng)
{
// check datatype
constexpr bool is_half = std::is_same<Y, half_t>::value;
constexpr bool is_float = std::is_same<Y, float>::value;
static_assert(is_half || is_float, "only half and float are supported.");
return run_cast_from_f8<X, Y, negative_zero_nan>(x);
return bit_cast<Y>(run_cast_to_f8<X, Y, clip, stoch>(x, rng));
}
} // namespace impl
CK_TILE_HOST_DEVICE fp8_raw_t float_to_fp8_sr_raw(float x)
#if CK_TILE_FP8_CVT_DEVICE
/**
* @brief Cast float to fp8/bf8 using device conversion instructions
*/
template <fp8_interpretation interpret, bool saturate, bool stochastic_rounding = false>
CK_TILE_DEVICE uint8_t cast_to_f8_from_f32(float v, unsigned int rng = 0)
{
constexpr int seed = 42;
uint32_t rng = prand_generator_t<float, seed>{}(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx94__)
float max_fp8 = 240.0f;
x = x > max_fp8 ? max_fp8 : (x < -max_fp8 ? -max_fp8 : x);
uint8_t i8data;
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
unsigned int i32val;
unsigned char i8val[4]; // NOTE: not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_sr_fp8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival;
return val.i8val[0]; // little endian
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr fp8_rounding_mode rm = fp8_rounding_mode::stochastic;
return bit_cast<fp8_raw_t>(impl::cast_to_f8<float,
fp8_t,
negative_zero_nan,
clip,
(rm == fp8_rounding_mode::stochastic)>(x, rng));
#endif
}
CK_TILE_HOST_DEVICE bf8_raw_t float_to_bf8_sr_raw(float x)
{
constexpr int seed = 42;
uint32_t rng = prand_generator_t<float, seed>{}(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx94__)
union
unsigned int ival = 0;
val.fval = v;
if constexpr(saturate)
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_sr_bf8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival;
return val.i8val[0]; // little endian
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr fp8_rounding_mode rm = fp8_rounding_mode::stochastic;
return bit_cast<bf8_raw_t>(impl::cast_to_f8<float,
bf8_t,
negative_zero_nan,
clip,
(rm == fp8_rounding_mode::stochastic)>(x, rng));
#endif
if constexpr(interpret == fp8_interpretation::E4M3_FNUZ)
{
if((val.i32val & 0x7F800000) != 0x7F800000)
{ /// propagate NAN/INF, no clipping
val.fval = __builtin_amdgcn_fmed3f(val.fval, 240.0, -240.0);
}
}
else if constexpr(interpret == fp8_interpretation::E4M3_OCP)
{ // OCP type
if((val.i32val & 0x7F800000) != 0x7F800000)
{ /// propagate NAN/INF, no clipping
val.fval = __builtin_amdgcn_fmed3f(val.fval, 448.0, -448.0);
}
}
else
{
if((val.i32val & 0x7F800000) != 0x7F800000)
{ /// propagate NAN/INF, no clipping
val.fval = __builtin_amdgcn_fmed3f(val.fval, 57344.0, -57344.0);
}
}
}
if constexpr(stochastic_rounding)
{
ival = (interpret == fp8_interpretation::E4M3_FNUZ) ||
(interpret == fp8_interpretation::E4M3_OCP)
? __builtin_amdgcn_cvt_sr_fp8_f32(val.fval, rng, ival, 0)
: __builtin_amdgcn_cvt_sr_bf8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival;
i8data = val.i8val[0]; // little endian
}
else
{ // RNE CVT
ival = (interpret == fp8_interpretation::E4M3_FNUZ) ||
(interpret == fp8_interpretation::E4M3_OCP)
? __builtin_amdgcn_cvt_pk_fp8_f32(val.fval, val.fval, ival, false)
: __builtin_amdgcn_cvt_pk_bf8_f32(val.fval,
val.fval,
ival,
false); // false -> WORD0
val.i32val = ival;
i8data = val.i8val[0];
}
return i8data;
}
#endif // CK_TILE_FP8_CVT_DEVICE
CK_TILE_HOST_DEVICE fp8_raw_t float_to_fp8_rtn_raw(float x)
} // namespace impl
/**
* @brief Converts a floating-point value to an 8-bit floating-point representation with stochastic
* rounding.
*
* This function converts a floating-point value (float or half_t) to an 8-bit floating-point
* representation of type fp8_t or bf8_t. The conversion process may
* involve clipping and uses a pseudo-random number generator for the stochastic rounding.
*
* @tparam DstT The destination type (fp8_t or bf8_t).
* @tparam SrcT The source type (float or half_t) to be converted.
* @param x The floating-point value to be converted.
* @return The 8-bit floating-point representation of the input value.
*/
template <typename SrcT, typename DstT>
CK_TILE_HOST_DEVICE typename numeric_traits<DstT>::bitwise_type float_to_fp8_sr_raw(SrcT x)
{
#if defined(__gfx94__)
float max_fp8 = 240.0f;
x = x > max_fp8 ? max_fp8 : (x < -max_fp8 ? -max_fp8 : x);
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_pk_fp8_f32(val.fval, val.fval, ival, false); // false -> WORD0
val.i32val = ival;
return val.i8val[0];
constexpr bool clip = true;
constexpr int seed = 42;
uint32_t rng = prand_generator_t<SrcT, seed>{}(reinterpret_cast<uintptr_t>(&x), x);
#if CK_TILE_FP8_CVT_DEVICE
return impl::cast_to_f8_from_f32<numeric_traits<DstT>::f8_interpret, clip, true>(x, rng);
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr fp8_rounding_mode rm = fp8_rounding_mode::standard;
constexpr uint32_t rng = 0;
return bit_cast<fp8_raw_t>(impl::cast_to_f8<float,
fp8_t,
negative_zero_nan,
clip,
(rm == fp8_rounding_mode::stochastic)>(x, rng));
return bit_cast<typename numeric_traits<DstT>::bitwise_type>(
impl::cast_to_f8<SrcT, DstT, clip, true>(x, rng));
#endif
}
CK_TILE_HOST_DEVICE bf8_raw_t float_to_bf8_rtn_raw(float x)
/**
* @brief Converts a floating-point value to an 8-bit floating-point representation with rounding to
* nearest even.
*
* This function converts a floating-point value (float or half_t) to an 8-bit floating-point
* representation of type fp8_t or bf8_t. The conversion process may involve clipping.
*
* @tparam DstT The destination type (fp8_t or bf8_t).
* @tparam SrcT The source type (float or half_t) to be converted.
* @param x The floating-point value to be converted.
* @return The 8-bit floating-point representation of the input value.
*/
template <typename SrcT, typename DstT>
CK_TILE_HOST_DEVICE typename numeric_traits<DstT>::bitwise_type float_to_fp8_rtn_raw(SrcT x)
{
#if defined(__gfx94__)
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_pk_bf8_f32(val.fval, val.fval, ival, false); // false -> WORD0
val.i32val = ival;
return val.i8val[0];
constexpr bool clip = true;
#if CK_TILE_FP8_CVT_DEVICE
return impl::cast_to_f8_from_f32<numeric_traits<DstT>::f8_interpret, clip, false>(x, 0);
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr fp8_rounding_mode rm = fp8_rounding_mode::standard;
constexpr uint32_t rng = 0;
return bit_cast<bf8_raw_t>(impl::cast_to_f8<float,
bf8_t,
negative_zero_nan,
clip,
(rm == fp8_rounding_mode::stochastic)>(x, rng));
return bit_cast<typename numeric_traits<DstT>::bitwise_type>(
impl::cast_to_f8<SrcT, DstT, clip, false>(x, 0));
#endif
}
// clang-format off
template<fp8_rounding_mode rounding>
template <fp8_rounding_mode rounding>
CK_TILE_HOST_DEVICE fp8_raw_t float_to_fp8_raw(float x, constant<rounding>)
{
if constexpr (rounding == fp8_rounding_mode::standard) return float_to_fp8_rtn_raw(x);
else if constexpr (rounding == fp8_rounding_mode::stochastic) return float_to_fp8_sr_raw(x);
else return fp8_raw_t{0};
if constexpr(rounding == fp8_rounding_mode::standard)
{
return float_to_fp8_rtn_raw<float, fp8_t>(x);
}
else if constexpr(rounding == fp8_rounding_mode::stochastic)
{
return float_to_fp8_sr_raw<float, fp8_t>(x);
}
else
{
return fp8_raw_t{0};
}
}
template<fp8_rounding_mode rounding>
template <fp8_rounding_mode rounding>
CK_TILE_HOST_DEVICE bf8_raw_t float_to_bf8_raw(float x, constant<rounding>)
{
if constexpr (rounding == fp8_rounding_mode::standard) return float_to_bf8_rtn_raw(x);
else if constexpr (rounding == fp8_rounding_mode::stochastic) return float_to_bf8_sr_raw(x);
else return bf8_raw_t{0};
if constexpr(rounding == fp8_rounding_mode::standard)
{
return float_to_fp8_rtn_raw<float, bf8_t>(x);
}
else if constexpr(rounding == fp8_rounding_mode::stochastic)
{
return float_to_fp8_sr_raw<float, bf8_t>(x);
}
else
{
return bf8_raw_t{0};
}
}
CK_TILE_HOST_DEVICE float fp8_to_float_raw(fp8_raw_t x)
{
#if defined(__gfx94__)
#if CK_TILE_FP8_CVT_DEVICE
float fval;
uint32_t i32val = static_cast<uint32_t>(x);
fval = __builtin_amdgcn_cvt_f32_fp8(i32val, 0);
// asm volatile("v_cvt_f32_fp8 %0, %1 src0_sel:BYTE_0" : "=v"(fval) : "v"(i32val));
return fval;
#else
constexpr bool negative_zero_nan = true;
return impl::cast_from_f8<fp8_t, float, negative_zero_nan>(bit_cast<fp8_t>(x));
return impl::run_cast_from_f8<fp8_t, float>(bit_cast<fp8_t>(x));
#endif
}
CK_TILE_HOST_DEVICE float bf8_to_float_raw(bf8_raw_t x)
{
#if defined(__gfx94__)
#if CK_TILE_FP8_CVT_DEVICE
float fval;
uint32_t i32val = static_cast<uint32_t>(x);
fval = __builtin_amdgcn_cvt_f32_bf8(i32val, 0);
// asm volatile("v_cvt_f32_bf8 %0, %1 src0_sel:BYTE_0" : "=v"(fval) : "v"(i32val));
return fval;
#else
constexpr bool negative_zero_nan = true;
return impl::cast_from_f8<bf8_t, float, negative_zero_nan>(bit_cast<bf8_t>(x));
return impl::run_cast_from_f8<bf8_t, float>(bit_cast<bf8_t>(x));
#endif
}
template<fp8_rounding_mode rounding = static_cast<fp8_rounding_mode>(CK_TILE_FLOAT_TO_FP8_DEFAULT)>
template <fp8_rounding_mode rounding = static_cast<fp8_rounding_mode>(CK_TILE_FLOAT_TO_FP8_DEFAULT)>
CK_TILE_HOST_DEVICE fp8_t float_to_fp8(float x, constant<rounding> = {})
{
return bit_cast<fp8_t>(float_to_fp8_raw(x, constant<rounding>{}));
}
template<fp8_rounding_mode rounding = static_cast<fp8_rounding_mode>(CK_TILE_FLOAT_TO_FP8_DEFAULT)>
template <fp8_rounding_mode rounding = static_cast<fp8_rounding_mode>(CK_TILE_FLOAT_TO_FP8_DEFAULT)>
CK_TILE_HOST_DEVICE bf8_t float_to_bf8(float x, constant<rounding> = {})
{
return bit_cast<bf8_t>(float_to_bf8_raw(x, constant<rounding>{}));
}
CK_TILE_HOST_DEVICE float fp8_to_float(fp8_t x)
{
return fp8_to_float_raw(bit_cast<fp8_raw_t>(x));
}
CK_TILE_HOST_DEVICE float fp8_to_float(fp8_t x) { return fp8_to_float_raw(bit_cast<fp8_raw_t>(x)); }
CK_TILE_HOST_DEVICE float bf8_to_float(bf8_t x)
{
return bf8_to_float_raw(bit_cast<bf8_raw_t>(x));
}
CK_TILE_HOST_DEVICE float bf8_to_float(bf8_t x) { return bf8_to_float_raw(bit_cast<bf8_raw_t>(x)); }
// clang-format on
template <typename T>
struct numeric_traits;
template <class T>
struct numeric;
#if CK_TILE_USE_OCP_FP8
template <>
struct numeric_traits<fp8_t>
struct numeric<fp8_t>
{
static constexpr int exp = 4;
static constexpr int mant = 3;
#if defined(__gfx94__)
static constexpr int bias = 8;
#else
static constexpr int bias = 7;
#endif
// minimum finite value, or minimum positive normal value
CK_TILE_HOST_DEVICE static constexpr fp8_t min()
{
return bit_cast<fp8_t>(static_cast<fp8_raw_t>(0x08)); // 0b00001000 = 2^-6
}
// minumum finite value
CK_TILE_HOST_DEVICE static constexpr fp8_t lowest()
{
return bit_cast<fp8_t>(static_cast<fp8_raw_t>(0xfe)); // 0b11111110 = -448
}
// maximum finite value
CK_TILE_HOST_DEVICE static constexpr fp8_t max()
{
return bit_cast<fp8_t>(static_cast<fp8_raw_t>(0x7e)); // 0b01111110 = 448
}
// difference between 1.0 and next representable f8 value (1.125)
// returns fp8_t(0.125)
CK_TILE_HOST_DEVICE static constexpr fp8_t epsilon()
{
return bit_cast<fp8_t>(static_cast<fp8_raw_t>(0x20)); // 0.125
}
// rounding error (0.0625)
// half of epsilon
CK_TILE_HOST_DEVICE static constexpr fp8_t round_error()
{
return bit_cast<fp8_t>(static_cast<fp8_raw_t>(0x18)); // 0.0625
}
// quiet NaN
CK_TILE_HOST_DEVICE static constexpr fp8_t quiet_NaN()
{
return bit_cast<fp8_t>(static_cast<fp8_raw_t>(0x7F)); // 0b01111111
}
// signaling NaN
CK_TILE_HOST_DEVICE static constexpr fp8_t signaling_NaN()
{
return bit_cast<fp8_t>(static_cast<fp8_raw_t>(0xFF)); // 0b11111111
}
// smallest positive subnormal value
CK_TILE_HOST_DEVICE static constexpr fp8_t denorm_min()
{
return bit_cast<fp8_t>(static_cast<fp8_raw_t>(0x01));
}
CK_TILE_HOST_DEVICE static constexpr fp8_t zero()
{
return bit_cast<fp8_t>(static_cast<fp8_raw_t>(0));
}
};
template <>
struct numeric_traits<bf8_t>
struct numeric<bf8_t>
{
static constexpr int exp = 5;
static constexpr int mant = 2;
#if defined(__gfx94__)
static constexpr int bias = 16;
#else
static constexpr int bias = 15; // IEEE
#endif
};
// minimum finite value, or minimum positive normalized value for float
CK_TILE_HOST_DEVICE static constexpr bf8_t min()
{
return bit_cast<bf8_t>(static_cast<bf8_raw_t>(0x04)); // 0b00000100 = 2^-14
}
template <class T>
struct numeric;
// minumum finite value
CK_TILE_HOST_DEVICE static constexpr bf8_t lowest()
{
return bit_cast<bf8_t>(static_cast<bf8_raw_t>(0xfb)); // 0b11111011 = -57344
}
// maximum finite value
CK_TILE_HOST_DEVICE static constexpr bf8_t max()
{
return bit_cast<bf8_t>(static_cast<bf8_raw_t>(0x7b)); // 0b01111011 = 57344
}
// difference between 1.0 and next representable bf8 value (1.25)
CK_TILE_HOST_DEVICE static constexpr bf8_t epsilon()
{
return bit_cast<bf8_t>(static_cast<bf8_raw_t>(0x34)); // 0.25
}
// rounding error (0.125)
// half of epsilon
CK_TILE_HOST_DEVICE static constexpr bf8_t round_error()
{
return bit_cast<bf8_t>(static_cast<bf8_raw_t>(0x30)); // 0.125
}
// positive infinity value
CK_TILE_HOST_DEVICE static constexpr bf8_t infinity()
{
return bit_cast<bf8_t>(static_cast<bf8_raw_t>(0x7c)); // 0b01111100
}
// quiet NaN
CK_TILE_HOST_DEVICE static constexpr bf8_t quiet_NaN()
{
return bit_cast<bf8_t>(static_cast<bf8_raw_t>(0x7F)); // 0b01111111
}
// signaling NaN
CK_TILE_HOST_DEVICE static constexpr bf8_t signaling_NaN()
{
return bit_cast<bf8_t>(static_cast<bf8_raw_t>(0xFF));
}
// smallest positive subnormal value
CK_TILE_HOST_DEVICE static constexpr bf8_t denorm_min()
{
return bit_cast<bf8_t>(static_cast<bf8_raw_t>(0x01));
}
CK_TILE_HOST_DEVICE static constexpr bf8_t zero()
{
return bit_cast<bf8_t>(static_cast<bf8_raw_t>(0));
}
};
#else
template <>
struct numeric<fp8_t>
{
......@@ -811,6 +1054,7 @@ struct numeric<bf8_t>
return bit_cast<bf8_t>(static_cast<bf8_raw_t>(0));
}
};
#endif
#if CK_TILE_USE_CUSTOM_DATA_TYPE
CK_TILE_ARITHMETIC_USING_FLOAT(CK_TILE_HOST_DEVICE, fp8_t)
......@@ -818,19 +1062,26 @@ CK_TILE_ARITHMETIC_USING_FLOAT(CK_TILE_HOST_DEVICE, bf8_t)
#endif
// math
CK_TILE_HOST_DEVICE
fp8_t abs(const fp8_t& x)
template <typename T>
CK_TILE_HOST_DEVICE T abs(const T& x)
{
return bit_cast<fp8_t>(static_cast<fp8_raw_t>(bit_cast<fp8_raw_t>(x) & 0x7f));
static_assert(std::is_same_v<T, fp8_t> || std::is_same_v<T, bf8_t>,
"Only fp8_t and bf8_t are supported");
return bit_cast<T>(static_cast<uint8_t>(bit_cast<uint8_t>(x) & numeric_traits<T>::abs_mask));
}
CK_TILE_HOST_DEVICE
bool isnan(const fp8_t& x)
{
uint8_t xx = bit_cast<fp8_raw_t>(x);
return xx == 0x80; // TODO: NANOO
}
#if CK_TILE_USE_OCP_FP8
return (xx & 0x7f) == 0x7f;
#else
return xx == 0x80;
#endif
}
#if CK_TILE_USE_CUSTOM_DATA_TYPE
CK_TILE_DEVICE
fp8_t sqrt(fp8_t x) { return static_cast<fp8_t>(__builtin_amdgcn_sqrtf(static_cast<float>(x))); };
......@@ -842,20 +1093,21 @@ fp8_t exp2(fp8_t x) { return static_cast<fp8_t>(exp2f(static_cast<float>(x))); }
CK_TILE_DEVICE
fp8_t log(fp8_t x) { return static_cast<fp8_t>(__logf(static_cast<float>(x))); };
CK_TILE_HOST_DEVICE
bf8_t abs(const bf8_t& x)
{
return bit_cast<bf8_t>(static_cast<fp8_raw_t>(bit_cast<bf8_raw_t>(x) & 0x7f));
}
#endif
CK_TILE_HOST_DEVICE
bool isnan(const bf8_t& x)
{
uint8_t xx = bit_cast<bf8_raw_t>(x);
return xx == 0x80; // TODO: NANOO
#if CK_TILE_USE_OCP_FP8
return (xx & 0x7f) > 0x7c;
#else
return xx == 0x80;
#endif
}
#if CK_TILE_USE_CUSTOM_DATA_TYPE
CK_TILE_DEVICE
bf8_t sqrt(bf8_t x) { return static_cast<bf8_t>(__builtin_amdgcn_sqrtf(static_cast<float>(x))); };
......@@ -867,5 +1119,6 @@ bf8_t exp2(bf8_t x) { return static_cast<bf8_t>(exp2f(static_cast<float>(x))); }
CK_TILE_DEVICE
bf8_t log(bf8_t x) { return static_cast<bf8_t>(__logf(static_cast<float>(x))); };
#endif
} // namespace ck_tile
include/ck_tile/core/numeric/half.hpp
View file @ 2a30cfdd
// 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.
#include "ck_tile/core/config.hpp"
#include "ck_tile/core/utility/bit_cast.hpp"
......@@ -236,10 +236,11 @@ struct numeric_traits<half_t>
static constexpr uint16_t head_mask = 0xFC00;
static constexpr uint16_t mant_mask = 0x3FF;
static constexpr uint16_t exp_mask = 0x1F;
static constexpr uint32_t Inf = 0x7C00;
static constexpr uint32_t NegInf = 0xFC00;
static constexpr uint32_t NaN = 0x7C01;
static constexpr uint32_t Neg0 = 0x8000;
static constexpr uint16_t abs_mask = 0x7FFF;
static constexpr uint16_t Inf = 0x7C00;
static constexpr uint16_t NegInf = 0xFC00;
static constexpr uint16_t NaN = 0x7C01;
static constexpr uint16_t Neg0 = 0x8000;
using bitwise_type = uint16_t;
};
......
include/ck_tile/core/numeric/numeric.hpp
View file @ 2a30cfdd
// 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
......@@ -89,6 +89,7 @@ struct numeric_traits<float>
static constexpr uint32_t head_mask = 0xFF800000;
static constexpr uint32_t mant_mask = 0x7FFFFF;
static constexpr uint32_t exp_mask = 0xFF;
static constexpr uint32_t abs_mask = 0x7FFFFFFF;
static constexpr uint32_t Inf = 0x7F800000;
static constexpr uint32_t NegInf = 0xFF800000;
static constexpr uint32_t NaN = 0x7F800001;
......
include/ck_tile/core/numeric/pk_int4.hpp 0 → 100644
View file @ 2a30cfdd
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck_tile/core/config.hpp"
#include "ck_tile/core/numeric/half.hpp"
#include "ck_tile/core/numeric/integral_constant.hpp"
#include "ck_tile/core/numeric/math.hpp"
#include "ck_tile/core/numeric/numeric.hpp"
#include "ck_tile/core/utility/bit_cast.hpp"
#include "ck_tile/core/utility/random.hpp"
#include <stdint.h>
#include <type_traits>
#include "ck_tile/core/numeric/int8.hpp"
#pragma once
namespace ck_tile {
// Packed 2xint4
struct pk_int4_t
{
using type = int8_t;
type data;
__host__ __device__ constexpr pk_int4_t() : data{type{}} {}
__host__ __device__ constexpr pk_int4_t(type init) : data{init} {}
};
// limits
template <class T>
struct numeric;
template <>
struct numeric<pk_int4_t>
{
// minimum finite value, or minimum positive normalized value for float
CK_TILE_HOST_DEVICE static constexpr pk_int4_t min()
{
constexpr uint8_t val = 0b10001000;
return pk_int4_t(bit_cast<int8_t>(val));
}
// minumum finite value
CK_TILE_HOST_DEVICE static constexpr pk_int4_t lowest()
{
constexpr uint8_t val = 0b10001000;
return pk_int4_t(bit_cast<int8_t>(val));
}
// maximum finite value
CK_TILE_HOST_DEVICE static constexpr pk_int4_t max()
{
constexpr uint8_t val = 0b01110111;
return pk_int4_t(bit_cast<int8_t>(val));
}
// difference between 1.0 and next value representable by float
CK_TILE_HOST_DEVICE static constexpr pk_int4_t epsilon()
{
return 1; // not used
}
CK_TILE_HOST_DEVICE static constexpr pk_int4_t round_error()
{
return 1; // not used
}
// positive infinity value
CK_TILE_HOST_DEVICE static constexpr pk_int4_t infinity()
{
return 1; // not used
}
// quiet NaN
CK_TILE_HOST_DEVICE static constexpr pk_int4_t quiet_NaN()
{
return 1; // not used
}
// signaling NaN
CK_TILE_HOST_DEVICE static constexpr pk_int4_t signaling_NaN()
{
return 1; // not used
}
// smallest positive subnormal value
CK_TILE_HOST_DEVICE static constexpr pk_int4_t denorm_min()
{
return 1; // not used
}
CK_TILE_HOST_DEVICE static constexpr pk_int4_t zero() { return 0; }
};
CK_TILE_HOST_DEVICE fp32x2_t pk_int4_t_to_fp32x2_t(const pk_int4_t& x)
{
uint8_t x_u8 = ck_tile::bit_cast<uint8_t>(x);
float x_l = ((x_u8 & 0x0f) >> 0) - 8.f;
float x_h = ((x_u8 & 0xf0) >> 4) - 8.f;
#ifdef CK_TILE_USE_PK4_LAYOUT_SHUFFLE
fp32x2_t res = {x_h, x_l};
#elif
fp32x2_t res = {x_l, x_h};
#endif
return res;
}
CK_TILE_HOST_DEVICE fp16x2_t pk_int4_t_to_halfx2_t(const pk_int4_t& x)
{
uint8_t x_u8 = ck_tile::bit_cast<uint8_t>(x);
#ifdef CK_TILE_USE_PK4_LAYOUT_SHUFFLE
uint32_t i4s = ((x_u8 & 0x0f) << 16) | ((x_u8 & 0xf0) >> 4);
#elif
uint32_t i4s = ((x_u8 & 0xf0) << 12) | (x_u8 & 0xf);
#endif
const int EX = 0x64006400;
const int SUB = 0xE408E408; //-8
int lo = i4s | EX;
return pk_add_f16(bit_cast<fp16x2_t>(lo), bit_cast<fp16x2_t>(SUB));
}
CK_TILE_HOST_DEVICE bf16x2_t pk_int4_t_to_bfloat16x2_t(const pk_int4_t& x)
{
uint8_t x_u8 = ck_tile::bit_cast<uint8_t>(x);
float x_l = ((x_u8 & 0x0f) >> 0) - 8.f;
float x_h = ((x_u8 & 0xf0) >> 4) - 8.f;
#ifdef CK_TILE_USE_PK4_LAYOUT_SHUFFLE
bf16x2_t res = {type_convert<bf16_t>(x_h), type_convert<bf16_t>(x_l)};
#elif
bf16x2_t res = {type_convert<bf16_t>(x_l), type_convert<bf16_t>(x_h)};
#endif
return res;
}
} // namespace ck_tile
include/ck_tile/core/numeric/vector_type.hpp
View file @ 2a30cfdd
// 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/static_distributed_tensor.hpp
View file @ 2a30cfdd
......@@ -29,6 +29,7 @@ struct static_distributed_tensor
remove_cvref_t<decltype(StaticTileDistribution{}.get_ys_to_d_descriptor())>;
static constexpr index_t kThreadElementSpaceSize = ThreadTensorDesc{}.get_element_space_size();
static_assert(0 < kThreadElementSpaceSize, "Make sure tile distribution is valid");
CK_TILE_HOST_DEVICE static constexpr auto get_num_of_dimension()
{
......
include/ck_tile/core/tensor/tile_window.hpp
View file @ 2a30cfdd
// 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 @ 2a30cfdd
// 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/type_traits.hpp
View file @ 2a30cfdd
......@@ -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 @ 2a30cfdd
......@@ -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()));
}
};
......
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