Skip to content

GitLab

Commit fdddc8f4 authored by Bartlomiej Kocot's avatar Bartlomiej Kocot
Browse files

Merge branch 'develop' of github.com:ROCmSoftwarePlatform/composable_kernel... 

Merge branch 'develop' of github.com:ROCmSoftwarePlatform/composable_kernel into barkocot/fix-cmake-tensor-op-instance
parents 8f48018d f7331c60
Hide whitespace changes
Inline Side-by-side
Showing with 511 additions and 320 deletions
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
View file @ fdddc8f4
......@@ -9,6 +9,7 @@
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor/static_tensor.hpp"
#include "ck/utility/is_detected.hpp"
namespace ck {
......@@ -211,10 +212,44 @@ struct ThreadwiseTensorSliceTransfer_v3r1
auto src_vector_container = src_vector_type{
src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
using dst_vector_type = vector_type_maker_t<DstData, SrcScalarPerVector>;
using dst_vector_t = typename dst_vector_type::type;
dst_vector_type op_r_v;
constexpr auto get_elem_op_vec_len = []() {
if constexpr(is_detected<is_pack8_invocable_t, decltype(src_element_op_)>::value)
{
if constexpr(decltype(src_element_op_)::is_pack8_invocable)
return math::min(8, SrcScalarPerVector);
}
if constexpr(is_detected<is_pack4_invocable_t, decltype(src_element_op_)>::value)
{
if constexpr(decltype(src_element_op_)::is_pack4_invocable)
return math::min(4, SrcScalarPerVector);
}
if constexpr(is_detected<is_pack2_invocable_t, decltype(src_element_op_)>::value)
{
if constexpr(decltype(src_element_op_)::is_pack2_invocable)
return math::min(2, SrcScalarPerVector);
}
return 1;
};
constexpr index_t elem_op_vec_len = get_elem_op_vec_len();
using src_elem_op_vec_t = typename vector_type<SrcData, elem_op_vec_len>::type;
using dst_elem_op_vec_t = typename vector_type<DstData, elem_op_vec_len>::type;
static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto idx) {
// apply the src elementwise op and convert to DstData under the hood if needed
src_element_op_(op_r_v.template AsType<dst_elem_op_vec_t>()(idx),
src_vector_container.template AsType<src_elem_op_vec_t>()[idx]);
});
// copy data from src_vector_container into src_thread_scratch_
src_thread_scratch_tuple_(thread_scratch_id)
.template SetAsType<src_vector_t>(
src_data_idx_seq, src_vector_container.template AsType<src_vector_t>()[I0]);
.template SetAsType<dst_vector_t>(src_data_idx_seq,
op_r_v.template AsType<dst_vector_t>()[I0]);
constexpr auto move_on_dim = [&]() constexpr
{
......@@ -267,19 +302,15 @@ struct ThreadwiseTensorSliceTransfer_v3r1
{
#if !CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE
static_ford<SliceLengths>{}([&](auto idx) {
// convert from SrcData to DstData here
dst_thread_scratch_(idx) =
type_convert<DstData>(src_thread_scratch_tuple_[thread_scratch_id][idx]);
dst_thread_scratch_(idx) = src_thread_scratch_tuple_[thread_scratch_id][idx];
});
#else
// sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_
// TODO make this logic more generic for more sub-dword datatype
if constexpr(SrcVectorDim != DstVectorDim &&
((is_same<half_t, remove_cvref_t<SrcData>>::value &&
is_same<half_t, remove_cvref_t<DstData>>::value &&
((is_same<half_t, remove_cvref_t<DstData>>::value &&
SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
(is_same<int8_t, remove_cvref_t<SrcData>>::value &&
is_same<int8_t, remove_cvref_t<DstData>>::value &&
(is_same<int8_t, remove_cvref_t<DstData>>::value &&
SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
{
// each transpose does
......@@ -313,7 +344,7 @@ struct ThreadwiseTensorSliceTransfer_v3r1
constexpr auto data_idx_seq = generate_sequence_v2(
[&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
using src_vector_t = vector_type_maker_t<SrcData, SrcScalarPerVector>;
using src_vector_t = vector_type_maker_t<DstData, SrcScalarPerVector>;
using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
// get DstScalarPerVector # of read-only references to src vectors from
......@@ -336,17 +367,16 @@ struct ThreadwiseTensorSliceTransfer_v3r1
Number<num_dst_vector>{});
// do data transpose
transpose_vectors<SrcData, DstScalarPerVector, SrcScalarPerVector>{}(
transpose_vectors<DstData, DstScalarPerVector, SrcScalarPerVector>{}(
src_vector_refs, dst_vector_refs);
});
}
static_ford<SliceLengths>{}([&](auto idx) {
// apply the src elementwise op and convert to DstData under the hood if needed
DstData dst_v;
src_element_op_(dst_v, src_thread_scratch_tuple_[thread_scratch_id][idx]);
dst_thread_scratch_(idx) = dst_v;
});
else
{
static_ford<SliceLengths>{}([&](auto idx) {
dst_thread_scratch_(idx) = src_thread_scratch_tuple_[thread_scratch_id][idx];
});
}
#endif
}
......@@ -761,11 +791,12 @@ struct ThreadwiseTensorSliceTransfer_v3r1
static constexpr auto src_thread_scratch_desc_ = decltype(GetSrcThreadScratchDescriptor()){};
static constexpr auto dst_thread_scratch_desc_ = decltype(GetDstThreadScratchDescriptor()){};
using SrcThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
SrcData,
SrcScalarPerVector,
decltype(src_thread_scratch_desc_),
true>;
using SrcThreadScratch =
StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
DstData, // apply data_convert with SrcThreadScratch
SrcScalarPerVector,
decltype(src_thread_scratch_desc_),
true>;
using DstThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
DstData,
......
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp
View file @ fdddc8f4
......@@ -132,9 +132,6 @@ struct ThreadwiseTensorSliceTransfer_v7r2
Number<num>{});
}
template <typename T>
using has_vec_len = decltype(std::declval<T&>().vec_len);
// SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
// SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
template <typename SrcBuffers,
......@@ -159,94 +156,63 @@ struct ThreadwiseTensorSliceTransfer_v7r2
is_src_valid);
});
if constexpr(is_detected<has_vec_len, decltype(element_op_)>::value)
{
constexpr auto elem_op_vec_len = decltype(element_op_)::vec_len;
static_assert(is_same<remove_cvref_t<decltype(elem_op_vec_len)>, index_t>::value,
"vec_len in element_op_ type is not index_t");
constexpr auto get_elem_op_vec_len = []() {
if constexpr(is_detected<is_pack8_invocable_t, decltype(element_op_)>::value)
{
if constexpr(decltype(element_op_)::is_pack8_invocable)
return math::min(8, SrcScalarPerVector);
}
if constexpr(is_detected<is_pack4_invocable_t, decltype(element_op_)>::value)
{
if constexpr(decltype(element_op_)::is_pack4_invocable)
return math::min(4, SrcScalarPerVector);
}
if constexpr(is_detected<is_pack2_invocable_t, decltype(element_op_)>::value)
{
if constexpr(decltype(element_op_)::is_pack2_invocable)
return math::min(2, SrcScalarPerVector);
}
return 1;
};
constexpr index_t elem_op_vec_len = get_elem_op_vec_len();
// apply pointwise function
static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto i) {
// get reference to src data
const auto src_data_refs = generate_tie(
// return type should be lvalue
[&](auto iSrc) -> const auto& {
using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
using elem_op_vec_t = typename vector_type<SrcData, elem_op_vec_len>::type;
return src_vectors[iSrc].template AsType<elem_op_vec_t>()[i];
},
Number<nSrc>{});
// get reference to dst data
auto dst_data_refs = generate_tie(
// return type should be lvalue
[&](auto iDst) -> auto& {
using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
using elem_op_vec_t = typename vector_type<DstData, elem_op_vec_len>::type;
return dst_vectors(iDst).template AsType<elem_op_vec_t>()(i);
},
Number<nDst>{});
static_assert(elem_op_vec_len == 1 || elem_op_vec_len == 2 ||
elem_op_vec_len == 4 || elem_op_vec_len == 8,
"vec_len in element_op_ must be 1, 2, 4, 8");
static_assert(SrcScalarPerVector % elem_op_vec_len == 0,
"vec_len in element_op_ cannot be divided by SrcScalarPerVector!");
// apply pointwise function
static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto i) {
// get reference to src data
const auto src_data_refs = generate_tie(
// return type should be lvalue
[&](auto iSrc) -> const auto& {
using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
using elem_op_vec_t =
typename vector_type<SrcData, elem_op_vec_len>::type;
return src_vectors[iSrc].template AsType<elem_op_vec_t>()[i];
},
Number<nSrc>{});
// get reference to dst data
auto dst_data_refs = generate_tie(
// return type should be lvalue
[&](auto iDst) -> auto& {
using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
using elem_op_vec_t =
typename vector_type<DstData, elem_op_vec_len>::type;
return dst_vectors(iDst).template AsType<elem_op_vec_t>()(i);
},
Number<nDst>{});
// apply pointwise function
// pointwise function signature:
// element_op_(dst_data_refs[I0],
// dst_data_refs[I1],
// ...,
// src_data_refs[I0],
// src_data_refs[I1],
// ...)
unpack2(element_op_, dst_data_refs, src_data_refs);
});
}
else
{
// apply pointwise function
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
// get reference to src data
const auto src_data_refs = generate_tie(
// return type should be lvalue
[&](auto iSrc) -> const auto& {
using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
return src_vectors[iSrc].template AsType<SrcData>()[i];
},
Number<nSrc>{});
// get reference to dst data
auto dst_data_refs = generate_tie(
// return type should be lvalue
[&](auto iDst) -> auto& {
using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
return dst_vectors(iDst).template AsType<DstData>()(i);
},
Number<nDst>{});
// apply pointwise function
// pointwise function signature:
// element_op_(dst_data_refs[I0],
// dst_data_refs[I1],
// ...,
// src_data_refs[I0],
// src_data_refs[I1],
// ...)
unpack2(element_op_, dst_data_refs, src_data_refs);
});
}
// pointwise function signature:
// element_op_(dst_data_refs[I0],
// dst_data_refs[I1],
// ...,
// src_data_refs[I0],
// src_data_refs[I1],
// ...)
unpack2(element_op_, dst_data_refs, src_data_refs);
});
dst_vectors_tuple_(iAccess) = dst_vectors;
......
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View file @ fdddc8f4
......@@ -462,7 +462,6 @@ struct mfma_type<MfmaInstr::mfma_f64_16x16x4f64>
}
};
#if defined CK_ENABLE_FP8
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16f8f8>
{
......@@ -506,9 +505,7 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32f8f8>
intrin_mfma_f32_16x16x32f8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
#endif
#if defined CK_ENABLE_BF8
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16bf8bf8>
{
......@@ -552,9 +549,7 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32bf8bf8>
intrin_mfma_f32_16x16x32bf8bf8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16f8bf8>
{
......@@ -598,9 +593,7 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32f8bf8>
intrin_mfma_f32_16x16x32f8bf8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16bf8f8>
{
......@@ -644,7 +637,6 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32bf8f8>
intrin_mfma_f32_16x16x32bf8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
#endif
template <typename base_type,
index_t MPerXdlops,
......@@ -792,7 +784,6 @@ struct MfmaSelector
}
#endif
#if defined CK_ENABLE_FP8
template <>
static constexpr auto GetMfma<f8_t, 32, 32>()
{
......@@ -804,9 +795,7 @@ struct MfmaSelector
{
return MfmaInstr::mfma_f32_16x16x32f8f8;
}
#endif
#if defined CK_ENABLE_BF8
template <>
static constexpr auto GetMfma<bf8_t, 32, 32>()
{
......@@ -818,9 +807,7 @@ struct MfmaSelector
{
return MfmaInstr::mfma_f32_16x16x32bf8bf8;
}
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
static constexpr auto GetMfma<f8_t, 32, 32, bf8_t>()
{
......@@ -832,9 +819,7 @@ struct MfmaSelector
{
return MfmaInstr::mfma_f32_16x16x32f8bf8;
}
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
static constexpr auto GetMfma<bf8_t, 32, 32, f8_t>()
{
......@@ -846,7 +831,6 @@ struct MfmaSelector
{
return MfmaInstr::mfma_f32_16x16x32bf8f8;
}
#endif
static constexpr auto selected_mfma =
mfma_type<GetMfma<base_type, MPerXdlops, NPerXdlops, additional_type>()>{};
......@@ -1051,18 +1035,10 @@ struct XdlopsGemm
static_assert(
is_same<base_type, double>::value || is_same<base_type, float>::value ||
is_same<base_type, half_t>::value || is_same<base_type, bhalf_t>::value ||
is_same<base_type, int8_t>::value
#if defined CK_ENABLE_FP8
|| is_same<base_type, f8_t>::value
#endif
#if defined CK_ENABLE_BF8
|| is_same<base_type, bf8_t>::value
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
|| (is_same<base_type, f8_t>::value && is_same<additional_type, bf8_t>::value) ||
(is_same<base_type, bf8_t>::value && is_same<additional_type, f8_t>::value)
#endif
,
is_same<base_type, int8_t>::value || is_same<base_type, f8_t>::value ||
is_same<base_type, bf8_t>::value ||
(is_same<base_type, f8_t>::value && is_same<additional_type, bf8_t>::value) ||
(is_same<base_type, bf8_t>::value && is_same<additional_type, f8_t>::value),
"base base_type must be double, float, half, bfloat16, int8_t, f8_t or bf8_t!");
static_for<0, KPack / mfma_instr.k_per_blk, 1>{}([&](auto k) {
......
include/ck/utility/amd_xdlops.hpp
View file @ fdddc8f4
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_AMD_XDLOPS_HPP
#define CK_AMD_XDLOPS_HPP
#include "data_type.hpp"
#pragma once
namespace ck {
......@@ -355,7 +352,6 @@ struct intrin_mfma_f64_16x16x4f64<16, 16>
}
};
#if defined CK_ENABLE_FP8
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16f8f8;
......@@ -418,9 +414,7 @@ struct intrin_mfma_f32_16x16x32f8f8<16, 16>
#endif
}
};
#endif
#if defined CK_ENABLE_BF8
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16bf8bf8;
......@@ -483,9 +477,7 @@ struct intrin_mfma_f32_16x16x32bf8bf8<16, 16>
#endif
}
};
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16f8bf8;
......@@ -548,9 +540,7 @@ struct intrin_mfma_f32_16x16x32f8bf8<16, 16>
#endif
}
};
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16bf8f8;
......@@ -613,6 +603,5 @@ struct intrin_mfma_f32_16x16x32bf8f8<16, 16>
#endif
}
};
#endif
} // namespace ck
#endif
include/ck/utility/data_type.hpp
View file @ fdddc8f4
......@@ -9,15 +9,9 @@ namespace ck {
using bhalf_t = ushort;
using half_t = _Float16;
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
using int4_t = _BitInt(4);
#endif
#if defined CK_ENABLE_FP8
using f8_t = _BitInt(8);
#endif
#if defined CK_ENABLE_BF8
using bf8_t = unsigned _BitInt(8);
#endif
using int4_t = _BitInt(4);
using f8_t = _BitInt(8);
using bf8_t = unsigned _BitInt(8);
// vector_type
template <typename T, index_t N>
......@@ -148,23 +142,19 @@ struct scalar_type<int4_t>
};
#endif
#if defined CK_ENABLE_FP8
template <>
struct scalar_type<f8_t>
{
using type = f8_t;
static constexpr index_t vector_size = 1;
};
#endif
#if defined CK_ENABLE_BF8
template <>
struct scalar_type<bf8_t>
{
using type = bf8_t;
static constexpr index_t vector_size = 1;
};
#endif
template <typename T>
struct vector_type<T, 1>
......@@ -968,24 +958,20 @@ using int8x32_t = typename vector_type<int8_t, 32>::type;
using int8x64_t = typename vector_type<int8_t, 64>::type;
// f8
#if defined CK_ENABLE_FP8
using f8x2_t = typename vector_type<f8_t, 2>::type;
using f8x4_t = typename vector_type<f8_t, 4>::type;
using f8x8_t = typename vector_type<f8_t, 8>::type;
using f8x16_t = typename vector_type<f8_t, 16>::type;
using f8x32_t = typename vector_type<f8_t, 32>::type;
using f8x64_t = typename vector_type<f8_t, 64>::type;
#endif
// bf8
#if defined CK_ENABLE_BF8
using bf8x2_t = typename vector_type<bf8_t, 2>::type;
using bf8x4_t = typename vector_type<bf8_t, 4>::type;
using bf8x8_t = typename vector_type<bf8_t, 8>::type;
using bf8x16_t = typename vector_type<bf8_t, 16>::type;
using bf8x32_t = typename vector_type<bf8_t, 32>::type;
using bf8x64_t = typename vector_type<bf8_t, 64>::type;
#endif
template <typename T>
struct NumericLimits
......@@ -1033,7 +1019,6 @@ struct NumericLimits<int4_t>
};
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
#if defined CK_ENABLE_FP8
template <>
struct NumericLimits<f8_t>
{
......@@ -1056,9 +1041,7 @@ struct NumericLimits<f8_t>
__host__ __device__ static constexpr f8_t QuietNaN() { return f8_t(binary_qnan); }
};
#endif
#if defined CK_ENABLE_BF8
template <>
struct NumericLimits<bf8_t>
{
......@@ -1081,7 +1064,6 @@ struct NumericLimits<bf8_t>
__host__ __device__ static constexpr bf8_t QuietNaN() { return bf8_t(binary_qnan); }
};
#endif
template <typename T>
struct NumericUtils
......@@ -1120,22 +1102,18 @@ struct NumericUtils<half_t>
using bitwise_type = uint16_t;
};
#if defined CK_ENABLE_FP8
template <>
struct NumericUtils<f8_t>
{
static constexpr int exp = 4;
static constexpr int mant = 3;
};
#endif
#if defined CK_ENABLE_BF8
template <>
struct NumericUtils<bf8_t>
{
static constexpr int exp = 5;
static constexpr int mant = 2;
};
#endif
//
} // namespace ck
include/ck/utility/f8_utils.hpp
View file @ fdddc8f4
......@@ -6,8 +6,6 @@
#include "ck/utility/data_type.hpp"
// these conversions are disabled if native conversions available
#if !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__)
#if defined CK_ENABLE_FP8 || defined CK_ENABLE_BF8
namespace ck {
// fp8 rounding modes
......@@ -244,5 +242,3 @@ __host__ __device__ Y cast_from_f8(X x)
}
} // namespace ck::utils
#endif // #if defined CK_ENABLE_FP8 || defined CK_ENABLE_BF8
#endif // #if !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__)
include/ck/utility/inner_product.hpp
View file @ fdddc8f4
......@@ -192,6 +192,8 @@ inner_product<int8x4_t, int8x4_t, int32_t>(const int8x4_t& a, const int8x4_t& b,
#else
c = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b), c, false);
#endif
#elif defined(CK_USE_AMD_V_DOT4_I32_I8_GFX11)
c = __builtin_amdgcn_sudot4(true, bit_cast<int32_t>(a), true, bit_cast<int32_t>(b), c, false);
#else
const vector_type<int8_t, 4> a_vector{a};
const vector_type<int8_t, 4> b_vector{b};
......
include/ck/utility/is_detected.hpp
View file @ fdddc8f4
......@@ -31,4 +31,13 @@ struct nonesuch
template <template <class...> class Op, class... Args>
using is_detected = typename detail::detector<nonesuch, void, Op, Args...>::value_t;
template <typename T>
using is_pack2_invocable_t = decltype(std::declval<T&>().is_pack2_invocable);
template <typename T>
using is_pack4_invocable_t = decltype(std::declval<T&>().is_pack4_invocable);
template <typename T>
using is_pack8_invocable_t = decltype(std::declval<T&>().is_pack8_invocable);
} // namespace ck
include/ck/utility/math.hpp
View file @ fdddc8f4
......@@ -150,28 +150,6 @@ __host__ __device__ constexpr T clamp(const T& x, const T& lowerbound, const T&
return min(max(x, lowerbound), upperbound);
}
// disallow implicit type casting
template <typename T>
__device__ T exp(T x);
// TODO: add f16 support using v_exp_f16
template <>
__device__ float exp<float>(float x)
{
return __expf(x);
}
template <>
__device__ double exp<double>(double x)
{
return exp(x);
}
static inline __host__ float exp(float x) { return std::expf(x); }
static inline __host__ double exp(double x) { return std::exp(x); }
// greatest common divisor, aka highest common factor
__host__ __device__ constexpr index_t gcd(index_t x, index_t y)
{
......
include/ck/utility/math_v2.hpp
View file @ fdddc8f4
......@@ -9,6 +9,7 @@
#include "ck/utility/data_type.hpp"
#include "ck/utility/type.hpp"
#include "ck/utility/type_convert.hpp"
namespace ck {
namespace math {
......@@ -92,14 +93,96 @@ static inline __host__ float sqrt(float x) { return std::sqrt(x); };
static inline __host__ double sqrt(double x) { return std::sqrt(x); };
static inline __host__ half_t tanh(half_t x)
template <typename T>
inline __host__ T tanh(T x)
{
return static_cast<half_t>(std::tanh(static_cast<float>(x)));
return ck::type_convert<T>(std::tanhf(ck::type_convert<float>(x)));
};
static inline __host__ float tanh(float x) { return std::tanh(x); };
template <>
inline __host__ float tanh<float>(float x)
{
return std::tanhf(x);
};
template <>
inline __host__ double tanh<double>(double x)
{
return std::tanh(x);
};
template <typename T>
inline __host__ T exp(T x)
{
return ck::type_convert<T>(std::expf(ck::type_convert<float>(x)));
}
template <>
inline __host__ float exp<float>(float x)
{
return std::expf(x);
}
static inline __host__ double tanh(double x) { return std::tanh(x); };
template <>
inline __host__ double exp<double>(double x)
{
return std::exp(x);
}
template <typename T>
inline __host__ T log(T x)
{
return ck::type_convert<T>(std::logf(ck::type_convert<float>(x)));
}
template <>
inline __host__ float log<float>(float x)
{
return std::logf(x);
}
template <>
inline __host__ double log<double>(double x)
{
return std::log(x);
}
template <typename T>
inline __host__ T pow(T x, T gamma)
{
return ck::type_convert<T>(
std::powf(ck::type_convert<float>(x), ck::type_convert<float>(gamma)));
}
template <>
inline __host__ float pow<float>(float x, float gamma)
{
return std::powf(x, gamma);
}
template <>
inline __host__ double pow<double>(double x, double gamma)
{
return std::pow(x, gamma);
}
template <typename T>
inline __host__ T expm1(T x)
{
return ck::type_convert<T>(std::expm1f(ck::type_convert<float>(x)));
}
template <>
inline __host__ float expm1<float>(float x)
{
return std::expm1f(x);
}
template <>
inline __host__ double expm1<double>(double x)
{
return std::expm1(x);
}
// math functions for the HIP kernel, some are implemented by calling hip builtin functions
......@@ -181,14 +264,107 @@ static inline __device__ float sqrt(float x) { return __builtin_amdgcn_sqrtf(x);
static inline __device__ double sqrt(double x) { return __builtin_amdgcn_sqrt(x); };
static inline __device__ half_t tanh(half_t x)
template <typename T>
inline __device__ T tanh(T x)
{
return ck::type_convert<T>(::tanhf(ck::type_convert<float>(x)));
};
template <>
inline __device__ float tanh<float>(float x)
{
return static_cast<half_t>(::tanhf(static_cast<float>(x)));
return ::tanhf(x);
};
static inline __device__ float tanh(float x) { return ::tanhf(x); };
template <>
inline __device__ double tanh<double>(double x)
{
return ::tanh(x);
};
template <typename T>
inline __device__ T exp(T x)
{
return ck::type_convert<T>(__expf(ck::type_convert<float>(x)));
};
template <>
inline __device__ half_t exp<half_t>(half_t x)
{
return hexp(x);
};
template <>
inline __device__ float exp<float>(float x)
{
return __expf(x);
};
static inline __device__ double tanh(double x) { return ::tanh(x); };
template <>
inline __device__ double exp<double>(double x)
{
return exp(x);
};
template <typename T>
inline __device__ T log(T x)
{
return ck::type_convert<T>(__logf(ck::type_convert<float>(x)));
};
template <>
inline __device__ half_t log<half_t>(half_t x)
{
return hlog(x);
};
template <>
inline __device__ float log<float>(float x)
{
return __logf(x);
};
template <>
inline __device__ double log<double>(double x)
{
return log(x);
};
template <typename T>
inline __device__ T pow(T x, T gamma)
{
return ck::type_convert<T>(powf(ck::type_convert<float>(x), ck::type_convert<float>(gamma)));
};
template <>
inline __device__ float pow<float>(float x, float gamma)
{
return powf(x, gamma);
};
template <>
inline __device__ double pow<double>(double x, double gamma)
{
return pow(x, gamma);
};
template <typename T>
inline __device__ T expm1(T x)
{
return ck::type_convert<T>(expm1f(ck::type_convert<float>(x)));
};
template <>
inline __device__ float expm1<float>(float x)
{
return expm1f(x);
};
template <>
inline __device__ double expm1<double>(double x)
{
return expm1(x);
};
} // namespace math
} // namespace ck
include/ck/utility/statically_indexed_array_multi_index.hpp
View file @ fdddc8f4
......@@ -5,6 +5,7 @@
#define CK_STATICALLY_INDEXED_ARRAY_MULTI_INDEX_HPP
#include "common_header.hpp"
#include "ck/utility/math_v2.hpp"
namespace ck {
......
include/ck/utility/type_convert.hpp
View file @ fdddc8f4
......@@ -95,7 +95,6 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_
return type_convert<bhalf_t>(x_fp32);
}
#if defined CK_ENABLE_FP8
// convert fp32 to fp8
template <>
inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
......@@ -146,7 +145,7 @@ inline __host__ __device__ f8_t type_convert<f8_t, half_t>(half_t x)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return type_convert<f8_t>(type_convert<float>(x));
#else
#elif 0
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
......@@ -154,6 +153,8 @@ inline __host__ __device__ f8_t type_convert<f8_t, half_t>(half_t x)
return utils::
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#else
return type_convert<f8_t>(type_convert<float>(x));
#endif
}
......@@ -164,14 +165,14 @@ inline __host__ __device__ half_t type_convert<half_t, f8_t>(f8_t x)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// use native conversion to float and convert to fp16
return type_convert<half_t>(type_convert<float>(x));
#else
#elif 0
constexpr bool negative_zero_nan = true;
return utils::cast_from_f8<f8_t, half_t, negative_zero_nan>(x);
#else
return type_convert<half_t>(type_convert<float>(x));
#endif
}
#endif
#if defined CK_ENABLE_BF8
// convert fp32 to bf8
template <>
inline __host__ __device__ bf8_t type_convert<bf8_t, float>(float x)
......@@ -222,7 +223,7 @@ inline __host__ __device__ bf8_t type_convert<bf8_t, half_t>(half_t x)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return type_convert<bf8_t>(type_convert<float>(x));
#else
#elif 0
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
......@@ -230,6 +231,8 @@ inline __host__ __device__ bf8_t type_convert<bf8_t, half_t>(half_t x)
return utils::
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#else
return type_convert<bf8_t>(type_convert<float>(x));
#endif
}
......@@ -240,12 +243,13 @@ inline __host__ __device__ half_t type_convert<half_t, bf8_t>(bf8_t x)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// use native conversion to float and convert to fp16
return type_convert<half_t>(type_convert<float>(x));
#else
#elif 0
constexpr bool negative_zero_nan = true;
return utils::cast_from_f8<bf8_t, half_t, negative_zero_nan>(x);
#else
return type_convert<half_t>(type_convert<float>(x));
#endif
}
#endif
// Declare a template function for bf16 conversion using RTN
template <typename Y, typename X>
......@@ -308,7 +312,6 @@ inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, half_t>(h
template <typename Y, typename X>
__host__ __device__ constexpr Y f8_convert_sr(X x);
#if defined CK_ENABLE_FP8
// convert fp32 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
......@@ -354,12 +357,10 @@ inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#else
return type_convert<f8_t>(type_convert<float>(x));
return f8_convert_sr<f8_t>(type_convert<float>(x));
#endif
}
#endif
#if defined CK_ENABLE_BF8
// convert fp32 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
......@@ -406,9 +407,8 @@ inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#else
return type_convert<bf8_t>(type_convert<float>(x));
return f8_convert_sr<bf8_t>(type_convert<float>(x));
#endif
}
#endif
} // namespace ck
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
View file @ fdddc8f4
......@@ -128,11 +128,9 @@ struct ReferenceConvFwd : public device::BaseOperator
}
}
float v_out;
arg.out_element_op_(v_out, v_acc);
arg.output_(g, n, k, wo) = ck::type_convert<OutDataType>(v_out);
OutDataType v_out;
arg.out_element_op_(v_out, ck::type_convert<OutDataType>(v_acc));
arg.output_(g, n, k, wo) = v_out;
};
make_ParallelTensorFunctor(func,
......@@ -184,11 +182,9 @@ struct ReferenceConvFwd : public device::BaseOperator
}
}
float v_out;
arg.out_element_op_(v_out, v_acc);
arg.output_(g, n, k, ho, wo) = ck::type_convert<OutDataType>(v_out);
OutDataType v_out;
arg.out_element_op_(v_out, ck::type_convert<OutDataType>(v_acc));
arg.output_(g, n, k, ho, wo) = v_out;
};
make_ParallelTensorFunctor(func,
......@@ -253,11 +249,9 @@ struct ReferenceConvFwd : public device::BaseOperator
}
}
float v_out;
arg.out_element_op_(v_out, v_acc);
arg.output_(g, n, k, d_o, ho, wo) = ck::type_convert<OutDataType>(v_out);
OutDataType v_out;
arg.out_element_op_(v_out, ck::type_convert<OutDataType>(v_acc));
arg.output_(g, n, k, d_o, ho, wo) = v_out;
};
make_ParallelTensorFunctor(func,
......
library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
View file @ fdddc8f4
......@@ -20,8 +20,9 @@ template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
typename AccDataType,
typename AccElementwiseOperation>
typename SaveMeanInvStdDataType,
typename ComputeDataType,
typename YElementwiseOperation>
struct ReferenceGroupnorm : public device::BaseOperator
{
// x = [N, H, W, G, C]
......@@ -35,14 +36,18 @@ struct ReferenceGroupnorm : public device::BaseOperator
const Tensor<GammaDataType>& gamma,
const Tensor<BetaDataType>& beta,
Tensor<YDataType>& y,
AccElementwiseOperation acc_elementwise_op,
Tensor<SaveMeanInvStdDataType>& save_mean,
Tensor<SaveMeanInvStdDataType>& save_inv_std,
YElementwiseOperation y_elementwise_op,
const std::vector<index_t> lengths,
AccDataType epsilon)
ComputeDataType epsilon)
: x_(x),
gamma_(gamma),
beta_(beta),
y_(y),
acc_elementwise_op_(acc_elementwise_op),
save_mean_(save_mean),
save_inv_std_(save_inv_std),
y_elementwise_op_(y_elementwise_op),
lengths_(lengths),
epsilon_(epsilon)
{
......@@ -52,9 +57,11 @@ struct ReferenceGroupnorm : public device::BaseOperator
const Tensor<XDataType> gamma_;
const Tensor<XDataType> beta_;
Tensor<YDataType>& y_;
AccElementwiseOperation acc_elementwise_op_;
Tensor<SaveMeanInvStdDataType>& save_mean_;
Tensor<SaveMeanInvStdDataType>& save_inv_std_;
YElementwiseOperation y_elementwise_op_;
std::vector<index_t> lengths_;
AccDataType epsilon_;
ComputeDataType epsilon_;
};
// Invoker
......@@ -68,8 +75,8 @@ struct ReferenceGroupnorm : public device::BaseOperator
int G = arg.lengths_[3];
int C = arg.lengths_[4];
Tensor<AccDataType> mean({N, G});
Tensor<AccDataType> var({N, G});
Tensor<ComputeDataType> mean({N, G});
Tensor<ComputeDataType> var({N, G});
// Compute mean & var in [H, W, C] by Welford Algorithm
// TODO - parallel for each HWC
......@@ -78,9 +85,9 @@ struct ReferenceGroupnorm : public device::BaseOperator
{
for(int g = 0; g < G; ++g)
{
AccDataType mean_val = type_convert<AccDataType>(0.0f);
AccDataType var_val = type_convert<AccDataType>(0.0f);
int32_t curr_count = 0;
ComputeDataType mean_val = type_convert<ComputeDataType>(0.0f);
ComputeDataType var_val = type_convert<ComputeDataType>(0.0f);
int32_t curr_count = 0;
for(int h = 0; h < H; ++h)
{
......@@ -89,10 +96,11 @@ struct ReferenceGroupnorm : public device::BaseOperator
for(int c = 0; c < C; ++c)
{
curr_count++;
AccDataType x = type_convert<AccDataType>(arg.x_(n, h, w, g, c));
AccDataType delta = x - mean_val;
ComputeDataType x =
type_convert<ComputeDataType>(arg.x_(n, h, w, g, c));
ComputeDataType delta = x - mean_val;
mean_val += delta / curr_count;
AccDataType delta2 = x - mean_val;
ComputeDataType delta2 = x - mean_val;
var_val += delta * delta2;
}
}
......@@ -100,6 +108,12 @@ struct ReferenceGroupnorm : public device::BaseOperator
mean(n, g) = mean_val;
var(n, g) = var_val / curr_count;
arg.save_mean_(n, g) = ck::type_convert<SaveMeanInvStdDataType>(mean(n, g));
ComputeDataType divisor =
static_cast<ComputeDataType>(1) / ck::math::sqrt(var(n, g) + arg.epsilon_);
arg.save_inv_std_(n, g) = ck::type_convert<SaveMeanInvStdDataType>(divisor);
}
}
......@@ -114,15 +128,19 @@ struct ReferenceGroupnorm : public device::BaseOperator
{
for(int c = 0; c < C; ++c)
{
AccDataType x = type_convert<AccDataType>(arg.x_(n, h, w, g, c));
AccDataType gamma = type_convert<AccDataType>(arg.gamma_(g, c));
AccDataType beta = type_convert<AccDataType>(arg.beta_(g, c));
AccDataType mean_val = type_convert<AccDataType>(mean(n, g));
AccDataType var_val = type_convert<AccDataType>(var(n, g));
AccDataType y = gamma * (x - mean_val) /
ck::math::sqrt(arg.epsilon_ + var_val) +
beta;
arg.acc_elementwise_op_(y, y);
ComputeDataType x =
type_convert<ComputeDataType>(arg.x_(n, h, w, g, c));
ComputeDataType gamma =
type_convert<ComputeDataType>(arg.gamma_(g, c));
ComputeDataType beta =
type_convert<ComputeDataType>(arg.beta_(g, c));
ComputeDataType mean_val =
type_convert<ComputeDataType>(mean(n, g));
ComputeDataType var_val = type_convert<ComputeDataType>(var(n, g));
ComputeDataType y = gamma * (x - mean_val) /
ck::math::sqrt(arg.epsilon_ + var_val) +
beta;
arg.y_elementwise_op_(y, y);
arg.y_(n, h, w, g, c) = type_convert<YDataType>(y);
}
}
......@@ -159,11 +177,14 @@ struct ReferenceGroupnorm : public device::BaseOperator
const Tensor<GammaDataType>& gamma,
const Tensor<BetaDataType>& beta,
Tensor<YDataType>& y,
AccElementwiseOperation acc_elementwise_op,
Tensor<SaveMeanInvStdDataType>& save_mean,
Tensor<SaveMeanInvStdDataType>& save_inv_std,
YElementwiseOperation y_elementwise_op,
const std::vector<index_t> lengths,
AccDataType epsilon)
ComputeDataType epsilon)
{
return Argument{x, gamma, beta, y, acc_elementwise_op, lengths, epsilon};
return Argument{
x, gamma, beta, y, save_mean, save_inv_std, y_elementwise_op, lengths, epsilon};
}
static auto MakeInvoker() { return Invoker{}; }
......
library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
View file @ fdddc8f4
......@@ -20,8 +20,9 @@ template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
typename AccDataType,
typename AccElementwiseOperation,
typename SaveMeanInvStdDataType,
typename ComputeDataType,
typename YElementwiseOperation,
index_t Rank,
index_t NumReduceDim>
struct ReferenceLayernorm : public device::BaseOperator
......@@ -36,15 +37,19 @@ struct ReferenceLayernorm : public device::BaseOperator
const Tensor<GammaDataType>& gamma_n,
const Tensor<BetaDataType>& beta_n,
Tensor<YDataType>& y_m_n,
AccElementwiseOperation acc_elementwise_op,
Tensor<SaveMeanInvStdDataType>& save_mean_m,
Tensor<SaveMeanInvStdDataType>& save_inv_std_m,
YElementwiseOperation y_elementwise_op,
const std::vector<index_t> lengths,
const std::vector<index_t> reduceDims,
AccDataType epsilon)
ComputeDataType epsilon)
: x_m_n_(x_m_n),
gamma_n_(gamma_n),
beta_n_(beta_n),
y_m_n_(y_m_n),
acc_elementwise_op_(acc_elementwise_op),
save_mean_m_(save_mean_m),
save_inv_std_m_(save_inv_std_m),
y_elementwise_op_(y_elementwise_op),
lengths_(lengths),
reduceDims_(reduceDims),
epsilon_(epsilon)
......@@ -55,10 +60,12 @@ struct ReferenceLayernorm : public device::BaseOperator
const Tensor<XDataType> gamma_n_;
const Tensor<XDataType> beta_n_;
Tensor<YDataType>& y_m_n_;
AccElementwiseOperation acc_elementwise_op_;
Tensor<SaveMeanInvStdDataType>& save_mean_m_;
Tensor<SaveMeanInvStdDataType>& save_inv_std_m_;
YElementwiseOperation y_elementwise_op_;
std::vector<index_t> lengths_;
std::vector<index_t> reduceDims_;
AccDataType epsilon_;
ComputeDataType epsilon_;
};
// Invoker
......@@ -69,8 +76,8 @@ struct ReferenceLayernorm : public device::BaseOperator
int M = arg.lengths_[0];
int N = arg.lengths_[1];
Tensor<AccDataType> mean({M});
Tensor<AccDataType> var({M});
Tensor<ComputeDataType> mean({M});
Tensor<ComputeDataType> var({M});
for(int m = 0; m < M; ++m)
{
......@@ -79,7 +86,7 @@ struct ReferenceLayernorm : public device::BaseOperator
for(int n = 0; n < N; ++n)
{
auto x_val = ck::type_convert<AccDataType>(arg.x_m_n_(m, n));
auto x_val = ck::type_convert<ComputeDataType>(arg.x_m_n_(m, n));
mean(m) += x_val;
var(m) += x_val * x_val;
}
......@@ -90,17 +97,21 @@ struct ReferenceLayernorm : public device::BaseOperator
for(int m = 0; m < M; ++m)
{
AccDataType divisor =
static_cast<AccDataType>(1) / ck::math::sqrt(var(m) + arg.epsilon_);
ComputeDataType divisor =
static_cast<ComputeDataType>(1) / ck::math::sqrt(var(m) + arg.epsilon_);
for(int n = 0; n < N; ++n)
{
auto x_val = ck::type_convert<AccDataType>(arg.x_m_n_(m, n));
auto y_val = (x_val - mean(m)) * divisor;
y_val = (y_val * arg.gamma_n_(n)) + arg.beta_n_(n);
arg.acc_elementwise_op_(y_val, y_val);
auto x_val = ck::type_convert<ComputeDataType>(arg.x_m_n_(m, n));
auto gamma_val = ck::type_convert<ComputeDataType>(arg.gamma_n_(n));
auto beta_val = ck::type_convert<ComputeDataType>(arg.beta_n_(n));
auto y_val = (x_val - mean(m)) * divisor;
y_val = (y_val * gamma_val) + beta_val;
arg.y_elementwise_op_(y_val, y_val);
arg.y_m_n_(m, n) = ck::type_convert<YDataType>(y_val);
}
arg.save_mean_m_(m) = ck::type_convert<SaveMeanInvStdDataType>(mean(m));
arg.save_inv_std_m_(m) = ck::type_convert<SaveMeanInvStdDataType>(divisor);
}
return 0;
......@@ -140,13 +151,23 @@ struct ReferenceLayernorm : public device::BaseOperator
const Tensor<GammaDataType>& gamma_n,
const Tensor<BetaDataType>& beta_n,
Tensor<YDataType>& y_m_n,
AccElementwiseOperation acc_elementwise_op,
Tensor<SaveMeanInvStdDataType>& save_mean_m,
Tensor<SaveMeanInvStdDataType>& save_inv_std_m,
YElementwiseOperation y_elementwise_op,
const std::vector<index_t> lengths,
const std::vector<index_t> reduceDims,
AccDataType epsilon)
ComputeDataType epsilon)
{
return Argument{
x_m_n, gamma_n, beta_n, y_m_n, acc_elementwise_op, lengths, reduceDims, epsilon};
return Argument{x_m_n,
gamma_n,
beta_n,
y_m_n,
save_mean_m,
save_inv_std_m,
y_elementwise_op,
lengths,
reduceDims,
epsilon};
}
static auto MakeInvoker() { return Invoker{}; }
......
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View file @ fdddc8f4
......@@ -20,12 +20,8 @@ using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using I32 = int32_t;
#if defined CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
#if defined CK_ENABLE_BF8
using BF8 = ck::bf8_t;
#endif
using F8 = ck::f8_t;
using BF8 = ck::bf8_t;
using Empty_Tuple = ck::Tuple<>;
......
library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
View file @ fdddc8f4
......@@ -240,11 +240,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, NWC> && is_same_v<WeiLayout, KXC> &&
is_same_v<OutLayout, NWK>)
{
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
......@@ -267,17 +269,23 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
}
#endif
}
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
{
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
}
#endif
#if defined(DL_KERNELS) && defined(CK_ENABLE_FP32)
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
......@@ -306,14 +314,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
}
#endif
}
else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
{
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
......
library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
View file @ fdddc8f4
......@@ -98,30 +98,31 @@ struct DeviceOperationInstanceFactory<
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
{
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
}
......
library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
View file @ fdddc8f4
......@@ -98,6 +98,26 @@ void add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);
#endif
template <typename ADataType,
......@@ -105,7 +125,8 @@ template <typename ADataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
typename CLayout,
typename ComputeType>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
BLayout,
......@@ -115,7 +136,8 @@ struct DeviceOperationInstanceFactory<
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>>
ck::tensor_operation::element_wise::PassThrough,
ComputeType>>
{
using DeviceOp = DeviceGemmSplitK<ALayout,
BLayout,
......@@ -125,14 +147,15 @@ struct DeviceOperationInstanceFactory<
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
ck::tensor_operation::element_wise::PassThrough,
ComputeType>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
is_same_v<CDataType, float>)
is_same_v<CDataType, float> && is_same_v<ComputeType, float>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -157,8 +180,8 @@ struct DeviceOperationInstanceFactory<
}
#endif
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t> && is_same_v<ComputeType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -183,8 +206,8 @@ struct DeviceOperationInstanceFactory<
}
#endif
#if(defined(CK_ENABLE_FP16) || defined(CK_ENABLE_FP8))
else if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t> && is_same_v<ComputeType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -207,8 +230,8 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instances(op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, half_t>)
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, half_t> && is_same_v<ComputeType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -231,6 +254,31 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t> && is_same_v<ComputeType, f8_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances(op_ptrs);
}
}
#endif
return op_ptrs;
}
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp
View file @ fdddc8f4
......@@ -6,8 +6,6 @@
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment