Skip to content

GitLab

Commit 9ba504b6 authored by ThomasNing's avatar ThomasNing
Browse files

merge with the develop support the fp8 with computev4

parents e3402c93 f49de496
Hide whitespace changes
Inline Side-by-side
Showing with 5218 additions and 70 deletions
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
View file @ 9ba504b6
......@@ -607,6 +607,7 @@ struct GridwiseBatchedGemmGemm_Xdl_CShuffle
// with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
// cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
// therefore we may just as well assign Gemm1KPack = group_size
constexpr index_t Gemm1KPack =
MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp
View file @ 9ba504b6
......@@ -856,11 +856,18 @@ struct GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
static_cast<A0B0B1DataType*>(p_shared) + SharedMemTrait::b1_block_space_offset,
b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr index_t Gemm1KPack = math::max(
math::lcm(
MfmaSelector<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl>::selected_mfma.group_size,
B1K1),
MfmaSelector<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl>::selected_mfma.k_per_blk);
// selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
// selected_mfma.k_per_blk <= Gemm1KPack
//
// Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
// multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
// Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
// with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
// cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
// therefore we may just as well assign Gemm1KPack = group_size
constexpr index_t Gemm1KPack =
MfmaSelector<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl>::selected_mfma.group_size;
auto blockwise_gemm1 = BlockwiseGemmXdlops_v2<
BlockSize,
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp
View file @ 9ba504b6
......@@ -773,6 +773,7 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle
// with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
// cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
// therefore we may just as well assign Gemm1KPack = group_size
constexpr index_t Gemm1KPack =
MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
View file @ 9ba504b6
......@@ -628,6 +628,7 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
// with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
// cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
// therefore we may just as well assign Gemm1KPack = group_size
constexpr index_t Gemm1KPack =
MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
......
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View file @ 9ba504b6
......@@ -37,7 +37,17 @@ enum struct MfmaInstr
mfma_f32_32x32x16f8bf8,
mfma_f32_16x16x32f8bf8,
mfma_f32_32x32x16bf8f8,
mfma_f32_16x16x32bf8f8
mfma_f32_16x16x32bf8f8,
mfma_f32_32x32x16f16,
mfma_f32_16x16x32f16,
mfma_f32_32x32x16bf16,
mfma_f32_16x16x32bf16,
mfma_i32_32x32x32i8,
mfma_i32_16x16x64i8,
mfma_f32_32x32x64f8f6f4,
mfma_f32_16x16x128f8f6f4,
mfma_scale_f32_32x32x64f8f6f4,
mfma_scale_f32_16x16x128f8f6f4
};
template <MfmaInstr instr>
......@@ -198,6 +208,50 @@ struct mfma_type<MfmaInstr::mfma_f32_32x32x8f16>
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16f16>
{
static constexpr index_t group_size = 4;
static constexpr index_t num_groups_per_blk = 4;
static constexpr index_t num_regs_per_blk = 16;
static constexpr index_t num_threads_per_blk = 32;
static constexpr index_t wave_size = 64;
static constexpr index_t num_input_blks = 2;
static constexpr index_t num_output_blks = 1;
static constexpr index_t m_per_blk = 32;
static constexpr index_t n_per_blk = 32;
static constexpr index_t k_per_blk = 8;
static constexpr bool is_k_reduction = true;
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_f32_32x32x16f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_16x16x32f16>
{
static constexpr index_t group_size = 4;
static constexpr index_t num_groups_per_blk = 1;
static constexpr index_t num_regs_per_blk = 4;
static constexpr index_t num_threads_per_blk = 16;
static constexpr index_t wave_size = 64;
static constexpr index_t num_input_blks = 4;
static constexpr index_t num_output_blks = 1;
static constexpr index_t m_per_blk = 16;
static constexpr index_t n_per_blk = 16;
static constexpr index_t k_per_blk = 8;
static constexpr bool is_k_reduction = true;
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_f32_16x16x32f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_16x16x16f16>
{
......@@ -264,6 +318,28 @@ struct mfma_type<MfmaInstr::mfma_f32_4x4x4f16>
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16bf16>
{
static constexpr index_t group_size = 4;
static constexpr index_t num_groups_per_blk = 4;
static constexpr index_t num_regs_per_blk = 16;
static constexpr index_t num_threads_per_blk = 32;
static constexpr index_t wave_size = 64;
static constexpr index_t num_input_blks = 2;
static constexpr index_t num_output_blks = 1;
static constexpr index_t m_per_blk = 32;
static constexpr index_t n_per_blk = 32;
static constexpr index_t k_per_blk = 8;
static constexpr bool is_k_reduction = true;
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_f32_32x32x16bf16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x8bf16_1k>
{
......@@ -286,6 +362,28 @@ struct mfma_type<MfmaInstr::mfma_f32_32x32x8bf16_1k>
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_16x16x32bf16>
{
static constexpr index_t group_size = 4;
static constexpr index_t num_groups_per_blk = 1;
static constexpr index_t num_regs_per_blk = 4;
static constexpr index_t num_threads_per_blk = 16;
static constexpr index_t wave_size = 64;
static constexpr index_t num_input_blks = 4;
static constexpr index_t num_output_blks = 1;
static constexpr index_t m_per_blk = 16;
static constexpr index_t n_per_blk = 16;
static constexpr index_t k_per_blk = 8;
static constexpr bool is_k_reduction = true;
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_f32_16x16x32bf16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_16x16x16bf16_1k>
{
......@@ -440,6 +538,50 @@ struct mfma_type<MfmaInstr::mfma_i32_16x16x32i8>
}
};
template <>
struct mfma_type<MfmaInstr::mfma_i32_32x32x32i8>
{
static constexpr index_t group_size = 4;
static constexpr index_t num_groups_per_blk = 4;
static constexpr index_t num_regs_per_blk = 16;
static constexpr index_t num_threads_per_blk = 32;
static constexpr index_t wave_size = 64;
static constexpr index_t num_input_blks = 2;
static constexpr index_t num_output_blks = 1;
static constexpr index_t m_per_blk = 32;
static constexpr index_t n_per_blk = 32;
static constexpr index_t k_per_blk = 16;
static constexpr bool is_k_reduction = true;
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_i32_32x32x32i8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_i32_16x16x64i8>
{
static constexpr index_t group_size = 4;
static constexpr index_t num_groups_per_blk = 1;
static constexpr index_t num_regs_per_blk = 4;
static constexpr index_t num_threads_per_blk = 16;
static constexpr index_t wave_size = 64;
static constexpr index_t num_input_blks = 4;
static constexpr index_t num_output_blks = 1;
static constexpr index_t m_per_blk = 16;
static constexpr index_t n_per_blk = 16;
static constexpr index_t k_per_blk = 16;
static constexpr bool is_k_reduction = true;
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_i32_16x16x64i8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f64_16x16x4f64>
{
......@@ -638,16 +780,115 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32bf8f8>
}
};
// TODO: fix mfma...f8f6f4 instructions
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x64f8f6f4>
{
// clang-format off
static constexpr index_t group_size = 4; // ??? group_size * num_groups_per_blk == num_regs_per_blk
static constexpr index_t num_groups_per_blk = 4; // ??? group_size * num_groups_per_blk == num_regs_per_blk
static constexpr index_t num_regs_per_blk = 16; // m_per_blk * n_per_blk / wave_size
static constexpr index_t num_threads_per_blk = 32; // n_per_blk
static constexpr index_t wave_size = 64; // fixed
static constexpr index_t num_input_blks = 2; // m_per_blk / num_regs_per_blk
static constexpr index_t num_output_blks = 1; // (is_k_reduction == true) ???
static constexpr index_t m_per_blk = 32; // from the instruction
static constexpr index_t n_per_blk = 32; // from the instruction
static constexpr index_t k_per_blk = 32; // (is_k_reduction == true) ? 64 / num_input_blks
static constexpr bool is_k_reduction = true; // ???
// clang-format on
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_f32_32x32x64f8f6f4<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_16x16x128f8f6f4>
{
// clang-format off
static constexpr index_t group_size = 4; // ??? group_size * num_groups_per_blk == num_regs_per_blk
static constexpr index_t num_groups_per_blk = 1; // ??? group_size * num_groups_per_blk == num_regs_per_blk
static constexpr index_t num_regs_per_blk = 4; // m_per_blk * n_per_blk / wave_size
static constexpr index_t num_threads_per_blk = 16; // == n_per_blk
static constexpr index_t wave_size = 64; // fixed
static constexpr index_t num_input_blks = 4; // m_per_blk / num_regs_per_blk
static constexpr index_t num_output_blks = 1; // (is_k_reduction == true) ???
static constexpr index_t m_per_blk = 16; // from the instruction
static constexpr index_t n_per_blk = 16; // from the instruction
static constexpr index_t k_per_blk = 32; // (is_k_reduction == true) ? 128 / num_input_blks
static constexpr bool is_k_reduction = true; // ???
// clang-format on
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_f32_16x16x128f8f6f4<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_scale_f32_32x32x64f8f6f4>
{
// clang-format off
static constexpr index_t group_size = 4; // ??? group_size * num_groups_per_blk == num_regs_per_blk
static constexpr index_t num_groups_per_blk = 4; // ??? group_size * num_groups_per_blk == num_regs_per_blk
static constexpr index_t num_regs_per_blk = 16; // m_per_blk * n_per_blk / wave_size
static constexpr index_t num_threads_per_blk = 32; // n_per_blk
static constexpr index_t wave_size = 64; // fixed
static constexpr index_t num_input_blks = 2; // m_per_blk / num_regs_per_blk
static constexpr index_t num_output_blks = 1; // (is_k_reduction == true) ???
static constexpr index_t m_per_blk = 32; // from the instruction
static constexpr index_t n_per_blk = 32; // from the instruction
static constexpr index_t k_per_blk = 32; // (is_k_reduction == true) ? 64 / num_input_blks
static constexpr bool is_k_reduction = true; // ???
// clang-format on
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_scale_f32_32x32x64f8f6f4<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_scale_f32_16x16x128f8f6f4>
{
// clang-format off
static constexpr index_t group_size = 4; // ??? group_size * num_groups_per_blk == num_regs_per_blk
static constexpr index_t num_groups_per_blk = 1; // ??? group_size * num_groups_per_blk == num_regs_per_blk
static constexpr index_t num_regs_per_blk = 4; // m_per_blk * n_per_blk / wave_size
static constexpr index_t num_threads_per_blk = 16; // == n_per_blk
static constexpr index_t wave_size = 64; // fixed
static constexpr index_t num_input_blks = 4; // m_per_blk / num_regs_per_blk
static constexpr index_t num_output_blks = 1; // (is_k_reduction == true) ???
static constexpr index_t m_per_blk = 16; // from the instruction
static constexpr index_t n_per_blk = 16; // from the instruction
static constexpr index_t k_per_blk = 32; // (is_k_reduction == true) ? 128 / num_input_blks
static constexpr bool is_k_reduction = true; // ???
// clang-format on
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_scale_f32_16x16x128f8f6f4<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <typename base_type,
index_t MPerXdlops,
index_t NPerXdlops,
typename additional_type = base_type>
typename additional_type = base_type,
bool is_single_rate_mfma = false>
struct MfmaSelector
{
template <typename base_type_,
index_t MPerXdlops_,
index_t NPerXdlops_,
typename additional_type_ = base_type_>
typename additional_type_ = base_type_,
bool is_single_rate_mfma_ = false>
static constexpr auto GetMfma();
template <>
......@@ -711,13 +952,32 @@ struct MfmaSelector
}
template <>
constexpr auto GetMfma<half_t, 32, 32>()
constexpr auto GetMfma<half_t, 32, 32, half_t, false>()
{
#if defined(__gfx950__)
return MfmaInstr::mfma_f32_32x32x16f16;
#else
return MfmaInstr::mfma_f32_32x32x8f16;
#endif
}
template <>
constexpr auto GetMfma<half_t, 32, 32, half_t, true>()
{
return MfmaInstr::mfma_f32_32x32x8f16;
}
template <>
constexpr auto GetMfma<half_t, 16, 16>()
constexpr auto GetMfma<half_t, 16, 16, half_t, false>()
{
#if defined(__gfx950__)
return MfmaInstr::mfma_f32_16x16x32f16;
#else
return MfmaInstr::mfma_f32_16x16x16f16;
#endif
}
template <>
constexpr auto GetMfma<half_t, 16, 16, half_t, true>()
{
return MfmaInstr::mfma_f32_16x16x16f16;
}
......@@ -741,7 +1001,19 @@ struct MfmaSelector
}
template <>
constexpr auto GetMfma<bhalf_t, 32, 32>()
constexpr auto GetMfma<bhalf_t, 32, 32, bhalf_t, false>()
{
#if defined(__gfx950__)
return MfmaInstr::mfma_f32_32x32x16bf16;
#elif defined(CK_USE_AMD_MFMA_BF16_1K_OP)
return MfmaInstr::mfma_f32_32x32x8bf16_1k;
#else
return MfmaInstr::mfma_f32_32x32x4bf16;
#endif
}
template <>
constexpr auto GetMfma<bhalf_t, 32, 32, bhalf_t, true>()
{
#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
return MfmaInstr::mfma_f32_32x32x8bf16_1k;
......@@ -751,7 +1023,19 @@ struct MfmaSelector
}
template <>
constexpr auto GetMfma<bhalf_t, 16, 16>()
constexpr auto GetMfma<bhalf_t, 16, 16, bhalf_t, false>()
{
#if defined(__gfx950__)
return MfmaInstr::mfma_f32_16x16x32bf16;
#elif defined(CK_USE_AMD_MFMA_BF16_1K_OP)
return MfmaInstr::mfma_f32_16x16x16bf16_1k;
#else
return MfmaInstr::mfma_f32_16x16x8bf16;
#endif
}
template <>
constexpr auto GetMfma<bhalf_t, 16, 16, bhalf_t, true>()
{
#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
return MfmaInstr::mfma_f32_16x16x16bf16_1k;
......@@ -760,7 +1044,18 @@ struct MfmaSelector
#endif
}
#if defined(CK_USE_AMD_MFMA_GFX940)
#if defined(__gfx950__)
template <>
constexpr auto GetMfma<int8_t, 32, 32>()
{
return MfmaInstr::mfma_i32_32x32x32i8;
}
template <>
constexpr auto GetMfma<int8_t, 16, 16>()
{
return MfmaInstr::mfma_i32_16x16x64i8;
}
#elif defined(__gfx942__)
template <>
constexpr auto GetMfma<int8_t, 32, 32>()
{
......@@ -832,8 +1127,8 @@ struct MfmaSelector
return MfmaInstr::mfma_f32_16x16x32bf8f8;
}
static constexpr auto selected_mfma =
mfma_type<GetMfma<base_type, MPerXdlops, NPerXdlops, additional_type>()>{};
static constexpr auto selected_mfma = mfma_type<
GetMfma<base_type, MPerXdlops, NPerXdlops, additional_type, is_single_rate_mfma>()>{};
__host__ __device__ constexpr MfmaSelector()
{
......@@ -1135,7 +1430,13 @@ struct XdlopsGemm
return TransposeC ? CIndex4D{blk_td, I0, blk_id, I0} : CIndex4D{I0, blk_id, I0, blk_td};
}
static constexpr auto mfma = MfmaSelector<base_type, MPerXdlops, NPerXdlops, additional_type>{};
// Falls back to single rate instruction on gfx950 if KPack <= 4; no change on gfx942-
static constexpr auto
mfma = MfmaSelector < base_type,
MPerXdlops, NPerXdlops, additional_type,
((is_same<base_type, half_t>::value || is_same<base_type, bhalf_t>::value) && KPack <= 4)
? true
: false > {};
static constexpr auto mfma_instr = mfma.selected_mfma;
......
include/ck/utility/amd_buffer_addressing.hpp
View file @ 9ba504b6
......@@ -581,7 +581,7 @@ __device__ void amd_global_atomic_add_impl(const typename vector_type<T, N>::typ
tmp.template AsType<half2_t>()[i]);
});
}
#if defined(__gfx942__)
#if defined(__gfx942__) || defined(__gfx950__)
else if constexpr(is_same<T, bhalf_t>::value)
{
vector_type<bhalf_t, N> tmp{src_thread_data};
......
include/ck/utility/amd_ck_fp8.hpp
View file @ 9ba504b6
......@@ -20,39 +20,25 @@
#define CK_USE_OCP_FP8 0
#endif
namespace {
// https://en.cppreference.com/w/cpp/types/conditional
template <bool B, class T, class F>
struct conditional
{
using type = T;
};
template <class T, class F>
struct conditional<false, T, F>
{
using type = F;
};
} // namespace
namespace ck {
using f8_fnuz_t = _BitInt(8);
using bf8_fnuz_t = unsigned _BitInt(8);
#if(defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx1200__) || \
defined(__gfx1201__)) && \
defined(__gfx1201__) || defined(__gfx950__)) && \
__HIP_DEVICE_COMPILE__
#define CK_FP8_CVT_FAST_PATH 1
#else
#define CK_FP8_CVT_FAST_PATH 0
#endif
#if(defined(__gfx1200__) || defined(__gfx1201__)) && __HIP_DEVICE_COMPILE__
#if(defined(__gfx1200__) || defined(__gfx1201__) || defined(__gfx950__)) && __HIP_DEVICE_COMPILE__
#define CK_OCP_FP8_CVT_FAST_PATH 1
#else
#define CK_OCP_FP8_CVT_FAST_PATH 0
#endif
namespace ck {
using f8_fnuz_t = _BitInt(8);
using bf8_fnuz_t = unsigned _BitInt(8);
typedef unsigned char fp8_storage_t;
/**
......@@ -207,10 +193,11 @@ __host__ __device__ static inline T cast_from_f8(fp8_storage_t x)
}
}
typename conditional<
typename std::conditional<
sizeof(T) == 2,
unsigned short int,
typename conditional<sizeof(T) == 4, unsigned int, unsigned long long>::type>::type retval;
typename std::conditional<sizeof(T) == 4, unsigned int, unsigned long long>::type>::type
retval;
if constexpr(we == 5 && is_half && !is_fnuz)
{
......@@ -303,7 +290,6 @@ static __device__ float2_t cast_to_f32x2_from_f8x2(fp8x2_storage_t v)
return __builtin_amdgcn_cvt_pk_f32_bf8(i16val, false);
}
}
#endif
} // namespace fp8_impl
......@@ -378,7 +364,7 @@ struct bf8_ocp_t
__host__ explicit operator float() const
#endif
{
#if defined(__gfx1200__) || defined(__gfx1201__)
#if defined(__gfx950__) || defined(__gfx1200__) || defined(__gfx1201__)
return fp8_impl::cast_to_f32_from_f8<default_interpret>(this->data);
#else
return fp8_impl::cast_from_f8<float, wm, we, false>(
......@@ -392,7 +378,7 @@ struct bf8_ocp_t
__host__ explicit operator _Float16() const
#endif
{
#if defined(__gfx1200__) || defined(__gfx1201__)
#if defined(__gfx950__) || defined(__gfx1200__) || defined(__gfx1201__)
return static_cast<_Float16>(fp8_impl::cast_to_f32_from_f8<default_interpret>(this->data));
#else
return fp8_impl::cast_from_f8<_Float16, wm, we, false>(
......@@ -553,10 +539,10 @@ __host__ __device__ static inline fp8_storage_t cast_to_f8(T _x, unsigned int rn
constexpr int mfmt = (sizeof(T) == 8) ? 52 : ((sizeof(T) == 4) ? 23 : 10);
using T_bitwise = typename conditional<
using T_bitwise = typename std::conditional<
sizeof(T) == 2,
unsigned short int,
typename conditional<sizeof(T) == 4, unsigned int, unsigned long long>::type>::type;
typename std::conditional<sizeof(T) == 4, unsigned int, unsigned long long>::type>::type;
T_bitwise x_bitwise = bit_cast<T_bitwise>(_x);
unsigned long long x{x_bitwise};
......
include/ck/utility/amd_xdlops.hpp
View file @ 9ba504b6
......@@ -5,7 +5,7 @@
namespace ck {
// Define the common macro for MI300 models
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
#define __gfx94__
#endif
......@@ -134,6 +134,46 @@ struct intrin_mfma_f32_32x32x4f16<32, 64>
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16f16;
template <>
struct intrin_mfma_f32_32x32x16f16<32, 32>
{
template <class FloatC>
__device__ static void Run(const half8_t& reg_a, const half8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x16_f16(
reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
#else
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif // defined(__gfx950__)
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_16x16x32f16;
template <>
struct intrin_mfma_f32_16x16x32f16<16, 16>
{
template <class FloatC>
__device__ static void Run(const half8_t& reg_a, const half8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x32_f16(
reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
#else
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif // defined(__gfx950__)
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x8f16;
......@@ -204,6 +244,46 @@ struct intrin_mfma_f32_4x4x4f16<8, 64>
};
// bfp16
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16bf16;
template <>
struct intrin_mfma_f32_32x32x16bf16<32, 32>
{
template <class FloatC>
__device__ static void Run(const bhalf8_t& reg_a, const bhalf8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x16_bf16(
reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
#else
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif // defined(__gfx950__)
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_16x16x32bf16;
template <>
struct intrin_mfma_f32_16x16x32bf16<16, 16>
{
template <class FloatC>
__device__ static void Run(const bhalf8_t& reg_a, const bhalf8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x32_bf16(
reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
#else
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif // defined(__gfx950__)
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x8bf16_1k;
......@@ -298,6 +378,46 @@ struct intrin_mfma_i32_16x16x16i8<16, 16>
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_i32_32x32x32i8;
template <>
struct intrin_mfma_i32_32x32x32i8<32, 32>
{
template <class FloatC>
__device__ static void Run(const int8x16_t& reg_a, const int8x16_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
reg_c.template AsType<int32x16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_i32_32x32x32_i8(
reg_a, reg_b, reg_c.template AsType<int32x16_t>()[Number<0>{}], 0, 0, 0);
#else
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif // defined(__gfx950__)
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_i32_16x16x64i8;
template <>
struct intrin_mfma_i32_16x16x64i8<16, 16>
{
template <class FloatC>
__device__ static void Run(const int8x16_t& reg_a, const int8x16_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
reg_c.template AsType<int32x4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_i32_16x16x64_i8(
reg_a, reg_b, reg_c.template AsType<int32x4_t>()[Number<0>{}], 0, 0, 0);
#else
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif // defined(__gfx950__)
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_i32_32x32x16i8;
......@@ -356,6 +476,149 @@ struct intrin_mfma_f64_16x16x4f64<16, 16>
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x64f8f6f4;
/// @brief Performs a matrix fused multiply-accumulate operation on 32x32x64 submatrices for f8, f6,
/// and f4 data types.
///
/// @note Calls scaled version of the instruction as the original instruction is not supported in
/// the backend. That is the intended use. There is a backend optimization to select the unscaled
/// operation if the scale is 0.
template <>
struct intrin_mfma_f32_32x32x64f8f6f4<32, 32>
{
template <class FloatC>
__device__ static void Run(const f8x32_t& reg_a, const f8x32_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
reg_c.template AsType<float16_t>()(Number<0>{}) =
__builtin_amdgcn_mfma_scale_f32_32x32x64_f8f6f4(
reg_a,
reg_b,
reg_c.template AsType<float16_t>()[Number<0>{}],
0, // cbsz
0, // blgp
0,
0,
0,
0);
#else
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_scale_f32_32x32x64f8f6f4;
template <>
struct intrin_mfma_scale_f32_32x32x64f8f6f4<32, 32>
{
template <class FloatC>
__device__ static void Run(const f8x32_t& reg_a,
const int32_t scale_a,
const f8x32_t& reg_b,
const int32_t scale_b,
FloatC& reg_c)
{
#if defined(__gfx950__)
// https://github.com/ROCm/llvm-project/blob/656552edc693e2bb4abc9258399c39d190fce2b3/llvm/test/Verifier/AMDGPU/mfma-scale.ll#L10
reg_c.template AsType<float16_t>()(Number<0>{}) =
__builtin_amdgcn_mfma_scale_f32_32x32x64_f8f6f4(
reg_a,
reg_b,
reg_c.template AsType<float16_t>()[Number<0>{}],
0, // cbsz
0, // blgp
0, // { OPSEL_HI[0], OPSEL[0] }?
scale_a,
0, // { OPSEL_HI[1], OPSEL[1] }?
scale_b);
#else
ignore = reg_a;
ignore = scale_a;
ignore = reg_b;
ignore = scale_b;
ignore = reg_c;
#endif
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_scale_f32_16x16x128f8f6f4;
template <>
struct intrin_mfma_scale_f32_16x16x128f8f6f4<16, 16>
{
template <class FloatC>
__device__ static void Run(const f8x32_t& reg_a,
const int32_t scale_a,
const f8x32_t& reg_b,
const int32_t scale_b,
FloatC& reg_c)
{
#if defined(__gfx950__)
// https://github.com/ROCm/llvm-project/blob/656552edc693e2bb4abc9258399c39d190fce2b3/llvm/test/Verifier/AMDGPU/mfma-scale.ll#L10
reg_c.template AsType<float4_t>()(Number<0>{}) =
__builtin_amdgcn_mfma_scale_f32_16x16x128_f8f6f4(
reg_a,
reg_b,
reg_c.template AsType<float4_t>()[Number<0>{}],
0, // cbsz
0, // blgp
0, // { OPSEL_HI[0], OPSEL[0] }?
scale_a,
0, // { OPSEL_HI[1], OPSEL[1] }?
scale_b);
#else
ignore = reg_a;
ignore = scale_a;
ignore = reg_b;
ignore = scale_b;
ignore = reg_c;
#endif
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_16x16x128f8f6f4;
/// @brief Performs a matrix fused multiply-accumulate operation on 16x16x128 submatrices for f8f6f4
/// data types.
///
/// @note Calls scaled version of the instruction as the original instruction is not supported in
/// the backend. That is the intended use. There is a backend optimization to select the unscaled
/// operation if the scale is 0.
template <>
struct intrin_mfma_f32_16x16x128f8f6f4<16, 16>
{
template <class FloatC>
__device__ static void Run(const f8x32_t& reg_a, const f8x32_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
// https://github.com/ROCm/llvm-project/blob/656552edc693e2bb4abc9258399c39d190fce2b3/llvm/test/Verifier/AMDGPU/mfma-scale.ll#L10
reg_c.template AsType<float4_t>()(Number<0>{}) =
__builtin_amdgcn_mfma_scale_f32_16x16x128_f8f6f4(
reg_a,
reg_b,
reg_c.template AsType<float4_t>()[Number<0>{}],
0, // cbsz
0, // blgp
0,
0,
0,
0);
#else
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16f8f8;
......
include/ck/utility/blkgemmpipe_scheduler.hpp
View file @ 9ba504b6
......@@ -90,14 +90,22 @@ struct BlockwiseGemmXdlops_pipeline_hotloop_inst
KPerXDL);
printf(" A/B buffer load inst: %d, %d\n A/B LDS write inst: %d, %d\n A/B LDS read inst: "
"%d, %d\n C MFMA inst: %d\n",
"%d, %d\n C MFMA inst: %d\n"
"A/B LDS read width: %d, %d, A/B LDS write width: %d, %d, A/B buffer load width: "
"%d/ %d\n",
A_Buffer_Load_Inst_Num,
B_Buffer_Load_Inst_Num,
A_LDS_Write_Inst_Num,
B_LDS_Write_Inst_Num,
A_LDS_Read_Inst_Num,
B_LDS_Read_Inst_Num,
C_MFMA_Inst_Num);
C_MFMA_Inst_Num,
A_LDS_Read_Width,
B_LDS_Read_Width,
ALDSWriteWidth,
BLDSWriteWidth,
ABufferLoadWidth,
BBufferLoadWidth);
}
};
......
include/ck/utility/data_type.hpp
View file @ 9ba504b6
......@@ -4,6 +4,7 @@
#pragma once
#include "ck/utility/amd_ck_fp8.hpp"
#include "ck/utility/e8m0.hpp"
#include "ck/utility/statically_indexed_array.hpp"
#ifdef CK_CODE_GEN_RTC
using int8_t = signed char;
......@@ -23,6 +24,296 @@ using std::byte;
using bhalf_t = ushort;
using half_t = _Float16;
using int4_t = _BitInt(4);
using f4_t = unsigned _BitInt(4);
using f6_t = _BitInt(6); // e2m3 format
using bf6_t = unsigned _BitInt(6); // e3m2 format
struct f4x2_pk_t
{
using type = uint8_t;
type data;
f4x2_pk_t() : data{type{}} {}
f4x2_pk_t(type init) : data{init} {}
template <index_t I>
__host__ __device__ inline type unpack(Number<I>) const
{
static_assert(I < 2, "Index is out of range.");
if constexpr(I == 0)
return data & 0b00001111;
else
return (data >> 4);
}
__host__ __device__ inline type pack(const type x0, const type x1)
{
return (x1 << 4) | (x0 & 0b00001111);
}
};
struct f6x16_pk_t
{
// store 16 elements of f6_t in an array of 3 uint32_t
using element_type = uint32_t;
using type = StaticallyIndexedArray_v2<element_type, 3>;
type data;
typedef int8_t test_vec_t __attribute__((ext_vector_type(16)));
f6x16_pk_t() : data{type{}} {}
f6x16_pk_t(type init) : data{init} {}
template <index_t I>
__host__ __device__ inline f6_t unpack(Number<I>)
{
static_assert(I < 16, "Index out of range for 16 f6_t elements.");
constexpr int num_bits_elem = 6;
constexpr int num_bits_vec_elem = 32;
constexpr int vector_size = 3;
constexpr int bit_pos = I * num_bits_elem;
constexpr int arr_idx = bit_pos / num_bits_vec_elem;
constexpr int bit_offset = bit_pos % num_bits_vec_elem;
uint32_t bits = data.At(Number<arr_idx>{}) >> bit_offset;
constexpr int overhang = bit_offset + num_bits_elem - num_bits_vec_elem;
if constexpr(overhang > 0 && (arr_idx + 1) < vector_size)
{
bits |= (data.At(Number<arr_idx + 1>{}) & ((1u << overhang) - 1))
<< (num_bits_elem - overhang);
}
return static_cast<f6_t>(bits & 0x3F);
}
__host__ __device__ inline type pack(const test_vec_t& x)
{
type packed{};
// for each of the 16 f6_t values, place its 6 bits in the correct position
ck::static_for<0, 16, 1>{}([&](auto i) {
uint32_t bits = static_cast<uint32_t>(x[static_cast<int>(i)]) & 0x3F;
constexpr int num_bits_elem = 6;
constexpr int num_bits_vec_elem = 32;
constexpr int vector_size = 3;
constexpr int bit_pos = i * num_bits_elem;
constexpr int arr_index = bit_pos / num_bits_vec_elem;
constexpr int bit_offset = bit_pos % num_bits_vec_elem;
constexpr int overhang = bit_offset + num_bits_elem - num_bits_vec_elem;
uint32_t old_value = packed.At(Number<arr_index>{});
// insert bits into the current 32-bit block
old_value |= (bits << bit_offset);
packed.At(Number<arr_index>{}) = old_value;
// if it crosses into the next block, shift the remainder
if constexpr(overhang > 0 && (arr_index + 1) < vector_size)
{
uint32_t next_value = packed.At(Number<arr_index + 1>{});
next_value |= (bits >> (num_bits_elem - overhang));
packed.At(Number<arr_index + 1>{}) = next_value;
}
});
return packed;
}
};
struct f6x32_pk_t
{
// store 32 elements of f6_t in an array of 6 uint32_t
using element_type = uint32_t;
using type = StaticallyIndexedArray_v2<element_type, 6>;
type data;
typedef int8_t test_vec_t __attribute__((ext_vector_type(32)));
f6x32_pk_t() : data{type{}} {}
f6x32_pk_t(type init) : data{init} {}
template <index_t I>
__host__ __device__ inline f6_t unpack(Number<I>)
{
static_assert(I < 32, "Index out of range for 32 f6_t elements.");
constexpr int num_bits_elem = 6;
constexpr int num_bits_vec_elem = 32;
constexpr int vector_size = 6;
constexpr int bit_pos = I * num_bits_elem;
constexpr int arr_idx = bit_pos / num_bits_vec_elem;
constexpr int bit_offset = bit_pos % num_bits_vec_elem;
uint32_t bits = data.At(Number<arr_idx>{}) >> bit_offset;
constexpr int overhang = bit_offset + num_bits_elem - num_bits_vec_elem;
if constexpr(overhang > 0 && (arr_idx + 1) < vector_size)
{
bits |= (data.At(Number<arr_idx + 1>{}) & ((1u << overhang) - 1))
<< (num_bits_elem - overhang);
}
return static_cast<f6_t>(bits & 0x3F);
}
__host__ __device__ inline type pack(const test_vec_t& x)
{
type packed{};
// for each of the 32 f6_t values, place its 6 bits in the correct position
ck::static_for<0, 32, 1>{}([&](auto i) {
uint32_t bits = static_cast<uint32_t>(x[static_cast<int>(i)]) & 0x3F;
constexpr int num_bits_elem = 6;
constexpr int num_bits_vec_elem = 32;
constexpr int vector_size = 6;
constexpr int bit_pos = i * num_bits_elem;
constexpr int arr_index = bit_pos / num_bits_vec_elem;
constexpr int bit_offset = bit_pos % num_bits_vec_elem;
constexpr int overhang = bit_offset + num_bits_elem - num_bits_vec_elem;
uint32_t old_value = packed.At(Number<arr_index>{});
// insert bits into the current 32-bit block
old_value |= (bits << bit_offset);
packed.At(Number<arr_index>{}) = old_value;
// if it crosses into the next block, shift the remainder
if constexpr(overhang > 0 && (arr_index + 1) < vector_size)
{
uint32_t next_value = packed.At(Number<arr_index + 1>{});
next_value |= (bits >> (num_bits_elem - overhang));
packed.At(Number<arr_index + 1>{}) = next_value;
}
});
return packed;
}
};
struct bf6x16_pk_t
{
// store 16 elements of bf6_t in an array of 3 uint32_t
using element_type = uint32_t;
using type = StaticallyIndexedArray_v2<element_type, 3>;
type data;
typedef int8_t test_vec_t __attribute__((ext_vector_type(16)));
bf6x16_pk_t() : data{type{}} {}
bf6x16_pk_t(type init) : data{init} {}
template <index_t I>
__host__ __device__ inline bf6_t unpack(Number<I>)
{
static_assert(I < 16, "Index out of range for 16 f6_t elements.");
constexpr int num_bits_elem = 6;
constexpr int num_bits_vec_elem = 32;
constexpr int vector_size = 3;
constexpr int bit_pos = I * num_bits_elem;
constexpr int arr_idx = bit_pos / num_bits_vec_elem;
constexpr int bit_offset = bit_pos % num_bits_vec_elem;
uint32_t bits = data.At(Number<arr_idx>{}) >> bit_offset;
constexpr int overhang = bit_offset + num_bits_elem - num_bits_vec_elem;
if constexpr(overhang > 0 && (arr_idx + 1) < vector_size)
{
bits |= (data.At(Number<arr_idx + 1>{}) & ((1u << overhang) - 1))
<< (num_bits_elem - overhang);
}
return static_cast<bf6_t>(bits & 0x3F);
}
__host__ __device__ inline type pack(const test_vec_t& x)
{
type packed{};
// for each of the 16 bf6_t values, place its 6 bits in the correct position
ck::static_for<0, 16, 1>{}([&](auto i) {
uint32_t bits = static_cast<uint32_t>(x[static_cast<int>(i)]) & 0x3F;
constexpr int num_bits_elem = 6;
constexpr int num_bits_vec_elem = 32;
constexpr int vector_size = 3;
constexpr int bit_pos = i * num_bits_elem;
constexpr int arr_index = bit_pos / num_bits_vec_elem;
constexpr int bit_offset = bit_pos % num_bits_vec_elem;
constexpr int overhang = bit_offset + num_bits_elem - num_bits_vec_elem;
uint32_t old_value = packed.At(Number<arr_index>{});
// insert bits into the current 32-bit block
old_value |= (bits << bit_offset);
packed.At(Number<arr_index>{}) = old_value;
// if it crosses into the next block, shift the remainder
if constexpr(overhang > 0 && (arr_index + 1) < vector_size)
{
uint32_t next_value = packed.At(Number<arr_index + 1>{});
next_value |= (bits >> (num_bits_elem - overhang));
packed.At(Number<arr_index + 1>{}) = next_value;
}
});
return packed;
}
};
struct bf6x32_pk_t
{
// store 32 elements of bf6_t in an array of 6 uint32_t
using element_type = uint32_t;
using type = StaticallyIndexedArray_v2<element_type, 6>;
type data;
typedef int8_t test_vec_t __attribute__((ext_vector_type(32)));
bf6x32_pk_t() : data{type{}} {}
bf6x32_pk_t(type init) : data{init} {}
template <index_t I>
__host__ __device__ inline bf6_t unpack(Number<I>)
{
static_assert(I < 32, "Index out of range for 32 f6_t elements.");
constexpr int num_bits_elem = 6;
constexpr int num_bits_vec_elem = 32;
constexpr int vector_size = 6;
constexpr int bit_pos = I * num_bits_elem;
constexpr int arr_idx = bit_pos / num_bits_vec_elem;
constexpr int bit_offset = bit_pos % num_bits_vec_elem;
uint32_t bits = data.At(Number<arr_idx>{}) >> bit_offset;
constexpr int overhang = bit_offset + num_bits_elem - num_bits_vec_elem;
if constexpr(overhang > 0 && (arr_idx + 1) < vector_size)
{
bits |= (data.At(Number<arr_idx + 1>{}) & ((1u << overhang) - 1))
<< (num_bits_elem - overhang);
}
return static_cast<bf6_t>(bits & 0x3F);
}
__host__ __device__ inline type pack(const test_vec_t& x)
{
type packed{};
// for each of the 32 bf6_t values, place its 6 bits in the correct position
ck::static_for<0, 32, 1>{}([&](auto i) {
uint32_t bits = static_cast<uint32_t>(x[static_cast<int>(i)]) & 0x3F;
constexpr int num_bits_elem = 6;
constexpr int num_bits_vec_elem = 32;
constexpr int vector_size = 6;
constexpr int bit_pos = i * num_bits_elem;
constexpr int arr_index = bit_pos / num_bits_vec_elem;
constexpr int bit_offset = bit_pos % num_bits_vec_elem;
constexpr int overhang = bit_offset + num_bits_elem - num_bits_vec_elem;
uint32_t old_value = packed.At(Number<arr_index>{});
// insert bits into the current 32-bit block
old_value |= (bits << bit_offset);
packed.At(Number<arr_index>{}) = old_value;
// if it crosses into the next block, shift the remainder
if constexpr(overhang > 0 && (arr_index + 1) < vector_size)
{
uint32_t next_value = packed.At(Number<arr_index + 1>{});
next_value |= (bits >> (num_bits_elem - overhang));
packed.At(Number<arr_index + 1>{}) = next_value;
}
});
return packed;
}
};
// custom data type - pack int4 data
struct pk_i4_t
......@@ -40,14 +331,15 @@ inline constexpr auto next_pow2(uint32_t x)
}
// native types: double, float, _Float16, ushort, int32_t, int8_t, uint8_t, f8_fnuz_t, bf8_fnuz_t,
// native types: bool
// native types: bool, f4_t, f6_t, bf6_t
template <typename T>
inline constexpr bool is_native_type()
{
return is_same<T, double>::value || is_same<T, float>::value || is_same<T, half_t>::value ||
is_same<T, bhalf_t>::value || is_same<T, int32_t>::value || is_same<T, int8_t>::value ||
is_same<T, uint8_t>::value || is_same<T, f8_fnuz_t>::value ||
is_same<T, bf8_fnuz_t>::value || is_same<T, bool>::value;
is_same<T, bf8_fnuz_t>::value || is_same<T, bool>::value || is_same<T, f4_t>::value ||
is_same<T, f6_t>::value || is_same<T, bf6_t>::value;
}
// vector_type
......@@ -1370,12 +1662,37 @@ struct nnvb_data_t_selector<f8_ocp_t>
{
using type = f8_ocp_t::data_type;
};
template <>
struct nnvb_data_t_selector<bf8_ocp_t>
{
using type = bf8_ocp_t::data_type;
};
template <>
struct nnvb_data_t_selector<f6x16_pk_t>
{
using type = f6x16_pk_t::type;
};
template <>
struct nnvb_data_t_selector<f6x32_pk_t>
{
using type = f6x32_pk_t::type;
};
template <>
struct nnvb_data_t_selector<bf6x16_pk_t>
{
using type = bf6x16_pk_t::type;
};
template <>
struct nnvb_data_t_selector<bf6x32_pk_t>
{
using type = bf6x32_pk_t::type;
};
template <>
struct nnvb_data_t_selector<pk_i4_t>
{
......@@ -1482,6 +1799,63 @@ struct non_native_vector_base<
}
};
// implementation for f6x16 and f6x32
template <typename T, index_t N>
struct non_native_vector_base<T, N, std::enable_if_t<sizeof(T) == 12 || sizeof(T) == 24>>
{
using data_t =
typename nnvb_data_t_selector<T>::type; // select data_t based on declared base type
using element_t = typename T::element_type; // select element_t based on declared element type
static_assert(sizeof(T) == sizeof(data_t), "non_native_vector_base storage size mismatch");
static constexpr size_t size_factor =
sizeof(data_t) / sizeof(element_t); // f6x16: 12/4 = 3, f6x32: 24/4 = 6
using data_v = element_t __attribute__((ext_vector_type(N * size_factor)));
using type = non_native_vector_base<T, N>;
union alignas(next_pow2(N * sizeof(T)))
{
data_v dN; // storage vector;
StaticallyIndexedArray<data_t, N> dxN;
StaticallyIndexedArray<T, N> dTxN;
StaticallyIndexedArray<data_v, 1> dNx1;
} data_;
__host__ __device__ constexpr non_native_vector_base(data_t a)
: data_{data_v(a.At(Number<0>{}))}
{
}
__host__ __device__ constexpr non_native_vector_base(T f)
: non_native_vector_base(bit_cast<data_t>(f))
{
}
__host__ __device__ constexpr non_native_vector_base() : non_native_vector_base(T{}){};
__host__ __device__ constexpr non_native_vector_base(data_v v) : data_{v} {}
__host__ __device__ constexpr operator data_v() const { return data_.dN; }
__host__ __device__ constexpr operator data_t() const
{
if constexpr(N == 1)
{
return data_.dxN[Number<0>{}];
}
else
{
return data_.dxN; // XXX this should cause an error
}
}
__host__ __device__ constexpr operator T() const
{
if constexpr(N == 1)
{
return data_.dTxN[Number<0>{}];
}
else
{
return data_.dTxN; // XXX this should cause an error
}
}
};
template <typename T, index_t N>
struct scalar_type<non_native_vector_base<T, N>>;
......@@ -2217,6 +2591,22 @@ using uint8x16_t = typename vector_type<uint8_t, 16>::type;
using uint8x32_t = typename vector_type<uint8_t, 32>::type;
using uint8x64_t = typename vector_type<uint8_t, 64>::type;
// f4
using f4x2_t = typename vector_type<f4x2_pk_t, 1>::type;
using f4x4_t = typename vector_type<f4x2_pk_t, 2>::type;
using f4x8_t = typename vector_type<f4x2_pk_t, 4>::type;
using f4x16_t = typename vector_type<f4x2_pk_t, 8>::type;
using f4x32_t = typename vector_type<f4x2_pk_t, 16>::type;
using f4x64_t = typename vector_type<f4x2_pk_t, 32>::type;
// f6
using f6x16_t = typename vector_type<f6x16_pk_t, 1>::type;
using f6x32_t = typename vector_type<f6x32_pk_t, 1>::type;
// bf6
using bf6x16_t = typename vector_type<bf6x16_pk_t, 1>::type;
using bf6x32_t = typename vector_type<bf6x32_pk_t, 1>::type;
// pack int4
using pk_i4x2_t = typename vector_type<pk_i4_t, 2>::type;
using pk_i4x4_t = typename vector_type<pk_i4_t, 4>::type;
......@@ -2571,6 +2961,118 @@ struct NumericLimits<bf8_ocp_t>
};
#endif
template <>
struct NumericLimits<f4_t>
{
static constexpr uint8_t binary_min_normal = 0x2; // 0b0010
static constexpr uint8_t binary_max_normal = 0x7; // 0b0111
static constexpr uint8_t binary_lowest_normal = 0xF; // 0b1111
static constexpr uint8_t binary_min_subnorm = 0x1; // 0b0001
static constexpr uint8_t binary_max_subnorm = 0x1; // 0b0001
static constexpr float data_max_normal_number = 6;
static constexpr float data_min_subnormal_number = 0.5;
__host__ __device__ static constexpr f4_t Min() { return f4_t(binary_min_normal); }
__host__ __device__ static constexpr f4_t Max() { return f4_t(binary_max_normal); }
__host__ __device__ static constexpr f4_t Lowest() { return f4_t(binary_lowest_normal); }
__host__ __device__ static constexpr f4_t MinSubnorm() { return f4_t(binary_min_subnorm); }
__host__ __device__ static constexpr f4_t MaxSubnorm() { return f4_t(binary_max_subnorm); }
__host__ __device__ static constexpr float DataMaxNorm() { return data_max_normal_number; }
__host__ __device__ static constexpr float DataMinSubnorm()
{
return data_min_subnormal_number;
}
};
template <>
struct NumericLimits<f6_t>
{
static constexpr uint8_t binary_min_normal = 0x08; // 0b001000
static constexpr uint8_t binary_max_normal = 0x1F; // 0b011111
static constexpr uint8_t binary_lowest_normal = 0x3F; // 0b111111
static constexpr uint8_t binary_min_subnorm = 0x01; // 0b000001
static constexpr uint8_t binary_max_subnorm = 0x07; // 0b000111
static constexpr float data_max_normal_number = 7.5;
static constexpr float data_min_subnormal_number = 0.125;
__host__ __device__ static constexpr f6_t Min() { return f6_t(binary_min_normal & 0b111111); }
__host__ __device__ static constexpr f6_t Max() { return f6_t(binary_max_normal & 0b111111); }
__host__ __device__ static constexpr f6_t Lowest()
{
return f6_t(binary_lowest_normal & 0b111111);
}
__host__ __device__ static constexpr f6_t MinSubnorm()
{
return f6_t(binary_min_subnorm & 0b111111);
}
__host__ __device__ static constexpr f6_t MaxSubnorm()
{
return f6_t(binary_max_subnorm & 0b111111);
}
__host__ __device__ static constexpr float DataMaxNorm() { return data_max_normal_number; }
__host__ __device__ static constexpr float DataMinSubnorm()
{
return data_min_subnormal_number;
}
};
template <>
struct NumericLimits<bf6_t>
{
static constexpr uint8_t binary_min_normal = 0x08; // 0b001000
static constexpr uint8_t binary_max_normal = 0x1F; // 0b011111
static constexpr uint8_t binary_lowest_normal = 0x3F; // 0b111111
static constexpr uint8_t binary_min_subnorm = 0x01; // 0b000001
static constexpr uint8_t binary_max_subnorm = 0x03; // 0b000011
static constexpr float data_max_normal_number = 28;
static constexpr float data_min_subnormal_number = 0.0625;
__host__ __device__ static constexpr bf6_t Min() { return bf6_t(binary_min_normal); }
__host__ __device__ static constexpr bf6_t Max() { return bf6_t(binary_max_normal); }
__host__ __device__ static constexpr bf6_t Lowest() { return bf6_t(binary_lowest_normal); }
__host__ __device__ static constexpr bf6_t MinSubnorm() { return bf6_t(binary_min_subnorm); }
__host__ __device__ static constexpr bf6_t MaxSubnorm() { return bf6_t(binary_max_subnorm); }
__host__ __device__ static constexpr float DataMaxNorm() { return data_max_normal_number; }
__host__ __device__ static constexpr float DataMinSubnorm()
{
return data_min_subnormal_number;
}
};
template <>
struct NumericLimits<e8m0_bexp_t>
{
static constexpr e8m0_bexp_t binary_min = 0x00; // 0b00000000
static constexpr e8m0_bexp_t binary_max = 0xFE; // 0b11111110
static constexpr e8m0_bexp_t binary_qnan = 0xFF; // 0b11111111
static constexpr e8m0_bexp_t binary_1 = 0x7F; // 0b01111111
static constexpr e8m0_bexp_t binary_2 = 0x80; // 0b10000000
static constexpr e8m0_bexp_t binary_3 = 0x82; // 0b10000010
static constexpr e8m0_bexp_t binary_135 = 0x87; // 0b10000111
static constexpr e8m0_bexp_t binary_142 = 0x8E; // 0b10001110
__host__ __device__ static constexpr e8m0_bexp_t Min() { return e8m0_bexp_t(binary_min); }
__host__ __device__ static constexpr e8m0_bexp_t Max() { return e8m0_bexp_t(binary_max); }
__host__ __device__ static constexpr e8m0_bexp_t QuietNaN() { return e8m0_bexp_t(binary_qnan); }
__host__ __device__ static constexpr e8m0_bexp_t Binary_1() { return e8m0_bexp_t(binary_1); }
__host__ __device__ static constexpr e8m0_bexp_t Binary_2() { return e8m0_bexp_t(binary_2); }
__host__ __device__ static constexpr e8m0_bexp_t Binary_3() { return e8m0_bexp_t(binary_3); }
__host__ __device__ static constexpr e8m0_bexp_t Binary_135()
{
return e8m0_bexp_t(binary_135);
}
__host__ __device__ static constexpr e8m0_bexp_t Binary_142()
{
return e8m0_bexp_t(binary_142);
}
};
template <typename T>
struct NumericUtils
{
......@@ -2590,6 +3092,7 @@ struct NumericUtils<float>
static constexpr uint32_t NegInf = 0xFF800000;
static constexpr uint32_t NaN = 0x7F800001;
static constexpr uint32_t Neg0 = 0x80000000;
static constexpr bool has_inf = true;
using bitwise_type = uint32_t;
};
......@@ -2607,9 +3110,19 @@ struct NumericUtils<half_t>
static constexpr uint32_t NegInf = 0xFC00;
static constexpr uint32_t NaN = 0x7C01;
static constexpr uint32_t Neg0 = 0x8000;
static constexpr bool has_inf = true;
using bitwise_type = uint16_t;
};
template <>
struct NumericUtils<bhalf_t>
{
static constexpr int exp = 8;
static constexpr int mant = 7;
static constexpr int bias = 128; // negative zero nan mode
// static constexpr int bias = 127; // ieee mode
};
template <>
struct NumericUtils<f8_fnuz_t>
{
......@@ -2617,6 +3130,7 @@ struct NumericUtils<f8_fnuz_t>
static constexpr int mant = 3;
static constexpr int bias = 8; // negative zero nan mode
// static constexpr int bias = 7; // ieee mode
static constexpr bool has_inf = false;
};
template <>
......@@ -2626,6 +3140,7 @@ struct NumericUtils<bf8_fnuz_t>
static constexpr int mant = 2;
static constexpr int bias = 16; // negative zero nan mode
// static constexpr int bias = 15; // ieee mode
static constexpr bool has_inf = false;
};
template <>
struct NumericUtils<f8_ocp_t>
......@@ -2644,11 +3159,109 @@ struct NumericUtils<bf8_ocp_t>
};
template <>
struct NumericUtils<bhalf_t>
struct NumericUtils<f4_t>
{
static constexpr int exp = 2;
static constexpr int mant = 1;
static constexpr int bias = 1;
static constexpr uint32_t sr_shift = 10;
static constexpr int unbiased_exp_min = 0;
static constexpr int unbiased_exp_max = 2;
static constexpr int biased_exp_min = 1;
static constexpr int biased_exp_max = 3;
static constexpr uint8_t positive_zero_mask = 0b0000;
static constexpr uint8_t negative_zero_mask = 0b1000;
static constexpr uint8_t one_mask = 0b0010;
static constexpr uint8_t set_sign_mask = 0b0111;
static constexpr uint8_t data_max_positive_normal_mask = 0b0111;
static constexpr uint8_t data_max_negative_normal_mask = 0b1111;
static constexpr uint8_t data_max_positive_subnormal_mask = 0b0001;
static constexpr uint8_t data_max_negative_subnormal_mask = 0b1001;
static constexpr bool has_inf = false;
using bitwise_type = uint8_t;
};
template <>
struct NumericUtils<f6_t>
{
static constexpr int exp = 2;
static constexpr int mant = 3;
static constexpr int bias = 1;
static constexpr uint32_t sr_shift = 12;
static constexpr int unbiased_exp_min = 0;
static constexpr int unbiased_exp_max = 2;
static constexpr int biased_exp_min = 1;
static constexpr int biased_exp_max = 3;
static constexpr uint8_t positive_zero_mask = 0b000000;
static constexpr uint8_t negative_zero_mask = 0b100000;
static constexpr uint8_t set_sign_mask = 0b011111;
static constexpr uint8_t data_max_positive_normal_mask = 0b011111;
static constexpr uint8_t data_max_negative_normal_mask = 0b111111;
static constexpr uint8_t data_max_positive_subnormal_mask = 0b000111;
static constexpr uint8_t data_max_negative_subnormal_mask = 0b100111;
static constexpr bool has_inf = false;
static constexpr bool has_nan = false;
static constexpr bool has_zero = true;
using bitwise_type = uint8_t;
};
template <>
struct NumericUtils<bf6_t>
{
static constexpr int exp = 3;
static constexpr int mant = 2;
static constexpr int bias = 3;
static constexpr uint32_t sr_shift = 11;
static constexpr int unbiased_exp_min = -2;
static constexpr int unbiased_exp_max = 4;
static constexpr int biased_exp_min = 1;
static constexpr int biased_exp_max = 7;
static constexpr uint8_t positive_zero_mask = 0b000000;
static constexpr uint8_t negative_zero_mask = 0b100000;
static constexpr uint8_t set_sign_mask = 0b011111;
static constexpr uint8_t data_max_positive_normal_mask = 0b011111;
static constexpr uint8_t data_max_negative_normal_mask = 0b111111;
static constexpr uint8_t data_max_positive_subnormal_mask = 0b000011;
static constexpr uint8_t data_max_negative_subnormal_mask = 0b100011;
static constexpr bool has_inf = false;
static constexpr bool has_nan = false;
static constexpr bool has_zero = true;
using bitwise_type = uint8_t;
};
template <>
struct NumericUtils<e8m0_bexp_t>
{
static constexpr int exp = 8;
static constexpr int mant = 7;
static constexpr int bias = 128; // negative zero nan mode
// static constexpr int bias = 127; // ieee mode
static constexpr int mant = 0;
static constexpr int bias = 127;
static constexpr int unbiased_exp_min = -127;
static constexpr int unbiased_exp_max = 127;
static constexpr int biased_exp_min = 0;
static constexpr int biased_exp_max = 254;
using bitwise_type = uint8_t;
};
} // namespace ck
include/ck/utility/e8m0.hpp 0 → 100644
View file @ 9ba504b6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/type.hpp"
namespace ck {
/**
* @brief Unsigned representation of a conventional biased Float32 exponent.
*
* bias = 127;
*
* E8M0_1 = 0b01111111; => 2^(127-127) = 1
* E8M0_2 = 0b10000000; => 2^(128-127) = 2^1 = 2
* E8M0_3 = 0b10000010; => 2^(130-127) = 2^3 = 8
* E8M0_135 = 0b10000111; => 2^(135-127) = 2^8 = 256
* E8M0_142 = 0b10001110; => 2^(142-127) = 2^15 = 32768
* E8M0_MIN = 0b00000000; => 2^-127
* E8M0_MAX = 0b11111110; => 2^127
* E8M0_NAN = 0b11111111; => NaN
*/
struct e8m0_bexp_t
{
using type = uint8_t;
type data;
constexpr static type bias = 127;
constexpr static type nan_mask = 0xFF;
__host__ __device__ constexpr e8m0_bexp_t() : data{type{}} {}
__host__ __device__ constexpr e8m0_bexp_t(type init) : data{init} {}
__host__ __device__ constexpr e8m0_bexp_t(int init) : data{static_cast<type>(init & nan_mask)}
{
}
__host__ __device__ explicit constexpr e8m0_bexp_t(float scale)
: data{static_cast<type>((bit_cast<uint32_t>(scale) & (nan_mask << 23)) >> 23)}
{
}
__host__ __device__ explicit constexpr operator float() const
{
if(data == nan_mask || data == 0)
{
uint32_t bits = data << 1;
bits |= 1;
bits <<= 22;
return bit_cast<float>(bits);
}
else
{
uint32_t bits = data << 23;
return bit_cast<float>(bits);
}
}
__host__ __device__ constexpr bool operator==(const e8m0_bexp_t& other) const
{
// strict IEEE compliance for NaN
return data == other.data && data != nan_mask;
}
__host__ __device__ constexpr bool is_nan() const { return data == nan_mask; }
};
namespace utils {
template <typename T>
__host__ __device__ inline int get_exponent_value(T x);
template <>
__host__ __device__ inline int get_exponent_value<e8m0_bexp_t>(e8m0_bexp_t x)
{
return x.data;
}
} // namespace utils
} // namespace ck
include/ck/utility/mxf4_utils.hpp 0 → 100644
View file @ 9ba504b6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/mxfp_utils.hpp"
namespace ck::utils {
template <>
__host__ __device__ inline bool is_nan<f4_t>(e8m0_bexp_t const scale,
f4_t const dataBytes [[maybe_unused]])
{
// no need to check for data as it does not have NaN representation
return scale == NumericLimits<e8m0_bexp_t>::QuietNaN();
}
// no infinity representation in ocp_e2m1_mxfp4 will always return false
template <>
__host__ __device__ inline bool is_inf<f4_t>(e8m0_bexp_t const scale [[maybe_unused]],
f4_t const data [[maybe_unused]])
{
// no inf representation for ocp_e2m1_mxfp4
return false;
}
template <>
__host__ __device__ inline bool is_zero<f4_t>(e8m0_bexp_t const scale, f4_t const data)
{
if(is_nan<f4_t>(scale, data))
return false;
// no need to check for scale as it does not have a 0 representation
f4_t result = (data & 0b00001111) & NumericUtils<f4_t>::set_sign_mask;
return result == 0b0;
}
template <>
__host__ __device__ inline float to_float<f4_t>(e8m0_bexp_t const scale, f4_t const data)
{
if(is_nan<f4_t>(scale, data))
return std::numeric_limits<float>::quiet_NaN();
if(is_zero<f4_t>(scale, data))
return 0.0f;
f4_t prepared_data = data & 0b00001111;
int scale_exp = get_exponent_value<e8m0_bexp_t>(scale);
return convert_to_float<f4_t>(prepared_data, scale_exp);
}
template <>
__host__ __device__ inline f4_t sat_convert_to_type<f4_t>(float value)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
{
return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
: NumericUtils<f4_t>::data_max_positive_normal_mask;
}
if(std::abs(value) > NumericLimits<f4_t>::Max()) // covers inf case as well
return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
: NumericUtils<f4_t>::data_max_positive_normal_mask;
f4_t res = convert_to_type<f4_t>(value);
if(std::abs(to_float<f4_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<f4_t>::DataMinSubnorm())
return value < 0 ? NumericUtils<f4_t>::negative_zero_mask
: NumericUtils<f4_t>::positive_zero_mask;
return res;
}
template <>
__host__ __device__ inline f4_t sat_convert_to_type_sr<f4_t>(float value, uint32_t seed)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
: NumericUtils<f4_t>::data_max_positive_normal_mask;
if(std::abs(value) > NumericLimits<f4_t>::Max()) // covers inf case as well
return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
: NumericUtils<f4_t>::data_max_positive_normal_mask;
f4_t res = convert_to_type_sr<f4_t>(value, seed);
if(std::abs(to_float<f4_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<f4_t>::DataMinSubnorm())
return value < 0 ? NumericUtils<f4_t>::negative_zero_mask
: NumericUtils<f4_t>::positive_zero_mask;
return res;
}
} // namespace ck::utils
include/ck/utility/mxf6_utils.hpp 0 → 100644
View file @ 9ba504b6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/mxfp_utils.hpp"
namespace ck::utils {
/**
* @brief Checks if an f6_t value is NaN based on the provided scale.
*
* For f6_t data, NaN cannot be represented directly. Instead, this function
* determines NaN by checking if the scale is set to a quiet NaN.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param dataBytes The f6_t value to check (unused in this implementation).
* @return true if the scale indicates a NaN value, false otherwise.
*/
template <>
__host__ __device__ inline bool is_nan<f6_t>(e8m0_bexp_t const scale,
f6_t const dataBytes [[maybe_unused]])
{
// no need to check for data as it does not have NaN representation
return scale.is_nan();
}
/**
* @brief Checks if an bf6_t value is NaN based on the provided scale.
*
* For bf6_t data, NaN cannot be represented directly. Instead, this function
* determines NaN by checking if the scale is set to a quiet NaN.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param dataBytes The bf6_t value to check (unused in this implementation).
* @return true if the scale indicates a NaN value, false otherwise.
*/
template <>
__host__ __device__ inline bool is_nan<bf6_t>(e8m0_bexp_t const scale,
bf6_t const dataBytes [[maybe_unused]])
{
// no need to check for data as it does not have NaN representation
return scale.is_nan();
}
/**
* @brief Checks if an f6_t value is infinite.
*
* Because f6_t does not support infinite values, this function always returns false.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param data The f6_t value to check.
* @return Always false, as infinity is not represented in f6_t.
*/
template <>
__host__ __device__ inline bool is_inf<f6_t>(e8m0_bexp_t const scale [[maybe_unused]],
f6_t const data [[maybe_unused]])
{
// no inf representation for fp6
return false;
}
/**
* @brief Checks if an bf6_t value is infinite.
*
* Because bf6_t does not support infinite values, this function always returns false.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param data The bf6_t value to check.
* @return Always false, as infinity is not represented in bf6_t.
*/
template <>
__host__ __device__ inline bool is_inf<bf6_t>(e8m0_bexp_t const scale [[maybe_unused]],
bf6_t const data [[maybe_unused]])
{
// no inf representation for bf6
return false;
}
/**
* @brief Checks whether an f6_t value is zero.
*
* If the specified f6_t is NaN, this function returns false.
* Otherwise, it masks out the sign bits and checks if the remaining bits
* are zero.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param data The f6_t value to check.
* @return true if the value is zero; otherwise false.
*/
template <>
__host__ __device__ inline bool is_zero<f6_t>(e8m0_bexp_t const scale, f6_t const data)
{
if(is_nan<f6_t>(scale, data))
return false;
// no need to check for scale as it does not have a 0 representation
f6_t result = (data & 0b00111111) & NumericUtils<f6_t>::set_sign_mask;
return result == 0b0;
}
/**
* @brief Checks whether an bf6_t value is zero.
*
* If the specified bf6_t is NaN, this function returns false.
* Otherwise, it masks out the sign bits and checks if the remaining bits
* are zero.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param data The bf6_t value to check.
* @return true if the value is zero; otherwise false.
*/
template <>
__host__ __device__ inline bool is_zero<bf6_t>(e8m0_bexp_t const scale, bf6_t const data)
{
if(is_nan<bf6_t>(scale, data))
return false;
// no need to check for scale as it does not have a 0 representation
bf6_t result = (data & 0b00111111) & NumericUtils<bf6_t>::set_sign_mask;
return result == 0b0;
}
/**
* @brief Converts an f6_t value to a float based on an e8m0_bexp_t scale factor.
*
* Checks if the f6_t value is NaN or zero before performing the conversion.
* Applies the exponent from the scale to compute the final float result.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param data The f6_t value to convert.
* @return The converted float value.
*/
template <>
__host__ __device__ inline float to_float<f6_t>(e8m0_bexp_t const scale, f6_t const data)
{
if(is_nan<f6_t>(scale, data))
return std::numeric_limits<float>::quiet_NaN();
if(is_zero<f6_t>(scale, data))
return 0.0f;
f6_t prepared_data = data & 0b00111111;
int scale_exp = get_exponent_value<e8m0_bexp_t>(scale);
return convert_to_float<f6_t>(prepared_data, scale_exp);
}
/**
* @brief Converts an bf6_t value to a float based on an e8m0_bexp_t scale factor.
*
* Checks if the bf6_t value is NaN or zero before performing the conversion.
* Applies the exponent from the scale to compute the final float result.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param data The bf6_t value to convert.
* @return The converted float value.
*/
template <>
__host__ __device__ inline float to_float<bf6_t>(e8m0_bexp_t const scale, bf6_t const data)
{
if(is_nan<bf6_t>(scale, data))
return std::numeric_limits<float>::quiet_NaN();
if(is_zero<bf6_t>(scale, data))
return 0.0f;
bf6_t prepared_data = data & 0b00111111;
int scale_exp = get_exponent_value<e8m0_bexp_t>(scale);
return convert_to_float<bf6_t>(prepared_data, scale_exp);
}
/**
* @brief Converts a float to f6_t with saturation.
*
* If the input is NaN or exceeds the representable range for f6_t, returns
* the corresponding max normal mask. Handles subnormal cases by returning
* zero with the appropriate sign.
*
* @param value The float value to be converted.
* @return The saturated f6_t value.
*/
template <>
__host__ __device__ inline f6_t sat_convert_to_type<f6_t>(float value)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
{
return sign ? NumericUtils<f6_t>::data_max_negative_normal_mask
: NumericUtils<f6_t>::data_max_positive_normal_mask;
}
if(std::abs(value) > NumericLimits<f6_t>::Max()) // covers inf case as well
return sign ? NumericUtils<f6_t>::data_max_negative_normal_mask
: NumericUtils<f6_t>::data_max_positive_normal_mask;
f6_t res = convert_to_type<f6_t>(value);
if(std::abs(to_float<f6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<f6_t>::DataMinSubnorm())
return sign ? NumericUtils<f6_t>::negative_zero_mask
: NumericUtils<f6_t>::positive_zero_mask;
return res;
}
/**
* @brief Converts a float to bf6_t with saturation.
*
* If the input is NaN or exceeds the representable range for bf6_t, returns
* the corresponding max normal mask. Handles subnormal cases by returning
* zero with the appropriate sign.
*
* @param value The float value to be converted.
* @return The saturated bf6_t value.
*/
template <>
__host__ __device__ inline bf6_t sat_convert_to_type<bf6_t>(float value)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
{
return sign ? NumericUtils<bf6_t>::data_max_negative_normal_mask
: NumericUtils<bf6_t>::data_max_positive_normal_mask;
}
if(std::abs(value) > NumericLimits<bf6_t>::Max()) // covers inf case as well
return sign ? NumericUtils<bf6_t>::data_max_negative_normal_mask
: NumericUtils<bf6_t>::data_max_positive_normal_mask;
bf6_t res = convert_to_type<bf6_t>(value);
if(std::abs(to_float<bf6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<bf6_t>::DataMinSubnorm())
return sign ? NumericUtils<bf6_t>::negative_zero_mask
: NumericUtils<bf6_t>::positive_zero_mask;
return res;
}
/**
* @brief Converts a float to f6_t with saturation and stochastic rounding.
*
* If the input is NaN or exceeds the representable range for f6_t, returns
* the corresponding max normal mask. Handles subnormal cases by returning
* zero with the appropriate sign.
*
* @param value The float value to be converted.
* @return The saturated f6_t value.
*/
template <>
__host__ __device__ inline f6_t sat_convert_to_type_sr<f6_t>(float value, uint32_t seed)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
return sign ? NumericUtils<f6_t>::data_max_negative_normal_mask
: NumericUtils<f6_t>::data_max_positive_normal_mask;
if(std::abs(value) > NumericLimits<f6_t>::Max()) // covers inf case as well
return sign ? NumericUtils<f6_t>::data_max_negative_normal_mask
: NumericUtils<f6_t>::data_max_positive_normal_mask;
f6_t res = convert_to_type_sr<f6_t>(value, seed);
if(std::abs(to_float<f6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<f6_t>::DataMinSubnorm())
return sign ? NumericUtils<f6_t>::negative_zero_mask
: NumericUtils<f6_t>::positive_zero_mask;
return res;
}
/**
* @brief Converts a float to f6_t with saturation and stochastic rounding.
*
* If the input is NaN or exceeds the representable range for f6_t, returns
* the corresponding max normal mask. Handles subnormal cases by returning
* zero with the appropriate sign.
*
* @param value The float value to be converted.
* @return The saturated f6_t value.
*/
template <>
__host__ __device__ inline bf6_t sat_convert_to_type_sr<bf6_t>(float value, uint32_t seed)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
return sign ? NumericUtils<bf6_t>::data_max_negative_normal_mask
: NumericUtils<bf6_t>::data_max_positive_normal_mask;
if(std::abs(value) > NumericLimits<bf6_t>::Max()) // covers inf case as well
return sign ? NumericUtils<bf6_t>::data_max_negative_normal_mask
: NumericUtils<bf6_t>::data_max_positive_normal_mask;
bf6_t res = convert_to_type_sr<bf6_t>(value, seed);
if(std::abs(to_float<bf6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<bf6_t>::DataMinSubnorm())
return sign ? NumericUtils<bf6_t>::negative_zero_mask
: NumericUtils<bf6_t>::positive_zero_mask;
return res;
}
} // namespace ck::utils
include/ck/utility/mxf8_utils.hpp 0 → 100644
View file @ 9ba504b6
#include "ck/utility/data_type.hpp"
#include "ck/utility/mxfp_utils.hpp"
#if defined(__gfx950__) && __HIP_DEVICE_COMPILE__
#define CK_MX_FP8_CVT_FAST_PATH 1
#else
#define CK_MX_FP8_CVT_FAST_PATH 0
#endif
namespace ck {
namespace fp8_impl {
#if CK_MX_FP8_CVT_FAST_PATH
template <ck_fp8_interpretation_t interpret>
static __device__ float cast_to_f32_from_f8_scaled(float scale, fp8_storage_t v)
{
union
{
unsigned int i32val;
unsigned char i8val[4];
} val;
val.i8val[0] = v;
static_assert(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP ||
interpret == ck_fp8_interpretation_t::CK_E5M2_OCP,
"Only OCP interpretations are supported");
if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
return __builtin_amdgcn_cvt_scalef32_f32_fp8(val.i32val, scale, 0);
}
else
{
return __builtin_amdgcn_cvt_scalef32_f32_bf8(val.i32val, scale, 0);
}
}
template <ck_fp8_interpretation_t interpret>
static __device__ float2_t cast_to_f32x2_from_f8x2_scaled(float scale, fp8x2_storage_t v)
{
const auto i16val = bit_cast<uint16_t>(v);
static_assert(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP ||
interpret == ck_fp8_interpretation_t::CK_E5M2_OCP,
"Only OCP interpretations are supported");
if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
return __builtin_amdgcn_cvt_scalef32_pk_f32_fp8(i16val, scale, 0);
}
else
{
return __builtin_amdgcn_cvt_scalef32_pk_f32_bf8(i16val, scale, 0);
}
}
template <ck_fp8_interpretation_t interpret, bool stochastic_rounding = false>
static __device__ fp8_storage_t cast_to_f8_from_f32_scaled(float v,
unsigned int rng = 0,
float scale = 1.0f)
{
fp8_storage_t i8data;
union
{
float fval;
unsigned int i32val;
} val;
union
{
uint32_t ival;
vector_type<int16_t, 2>::type v2i16;
fp8_storage_t v4i8[4];
} ret{};
// unsigned int ival = 0;
val.fval = v;
if constexpr(stochastic_rounding)
{
ret.ival =
(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
? __builtin_amdgcn_cvt_scalef32_sr_fp8_f32(ret.ival, val.fval, rng, scale, 0)
: __builtin_amdgcn_cvt_scalef32_sr_bf8_f32(ret.ival, val.fval, rng, scale, 0);
i8data = ret.v4i8[0];
}
else
{
// RNE CVT
// llvm.amdgcn.cvt.scalef32.pk.fp8.f32
// v2i16 old_vdst, float srcA, float srcB, float scale, bool dst_lo_hi_sel
if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
// If fval / scale > max fp8, returns Nan
ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_fp8_f32(/*old_vdst*/ ret.v2i16,
val.fval,
val.fval,
scale,
/*dst_lo_hi_sel*/ false);
}
else
{
// If fval / scale > max bf8, returns Inf
ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_bf8_f32(/*old_vdst*/ ret.v2i16,
val.fval,
val.fval,
scale,
/*dst_lo_hi_sel*/ false);
}
i8data = ret.v4i8[0];
}
return i8data;
}
template <ck_fp8_interpretation_t interpret, bool stochastic_rounding = false>
static __device__ fp8x2_storage_t cast_to_f8_from_f32_scaled(float2_t v,
unsigned int rng = 0,
float scale = 1.0f)
{
union
{
uint32_t ival;
vector_type<int16_t, 2>::type v2i16;
StaticallyIndexedArray<fp8x2_storage_t, 2> v2f8x2;
} ret{};
if constexpr(stochastic_rounding)
{
fp8x2_storage_t f8x2;
if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
ret.ival = __builtin_amdgcn_cvt_scalef32_sr_fp8_f32(ret.ival, v[0], rng, scale, 0);
f8x2[0] = ret.v2f8x2(Number<0>{})[0];
ret.ival = __builtin_amdgcn_cvt_scalef32_sr_fp8_f32(ret.ival, v[1], rng, scale, 0);
f8x2[1] = ret.v2f8x2(Number<0>{})[0];
}
else
{
ret.ival = __builtin_amdgcn_cvt_scalef32_sr_bf8_f32(ret.ival, v[0], rng, scale, 0);
f8x2[0] = ret.v2f8x2(Number<0>{})[0];
ret.ival = __builtin_amdgcn_cvt_scalef32_sr_bf8_f32(ret.ival, v[1], rng, scale, 0);
f8x2[1] = ret.v2f8x2(Number<0>{})[0];
}
return f8x2;
}
else
{
// RNE CVT
// llvm.amdgcn.cvt.scalef32.pk.fp8.f32
// v2i16 old_vdst, float srcA, float srcB, float scale, bool dst_lo_hi_sel
if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
// If fval / scale > max fp8, returns Nan
ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_fp8_f32(/*old_vdst*/ ret.v2i16,
v[0],
v[1],
scale,
/*dst_lo_hi_sel*/ false);
}
else
{
// If fval / scale > max bf8, returns Inf
ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_bf8_f32(/*old_vdst*/ ret.v2i16,
v[0],
v[1],
scale,
/*dst_lo_hi_sel*/ false);
}
return ret.v2f8x2(Number<0>{});
}
}
#endif // CK_MX_FP8_CVT_FAST_PATH
#if CK_MX_FP8_CVT_FAST_PATH
/**
* \brief convert float to @p fp8_storage_t with scaling
*
* This version is used when the fast path (MX FP8 hardware) is available
*
* \tparam interp interpretation of fp8
* \param f float number
* \param scale scaling factor
* \return fp8_storage_t
*/
template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
__host__ __device__ static inline fp8_storage_t cvt_float_to_fp8_scaled(const float f, float scale)
{
__is_interpret_supported(interp);
uint32_t rng = 0;
if constexpr(stochastic_rounding)
{
constexpr int seed = 1254739;
rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f);
}
return cast_to_f8_from_f32_scaled<interp, stochastic_rounding>(f, rng, scale);
}
/**
* \brief convert 2xfloat to @p 2xfp8_storage_t with scaling
*
* This version is used when the fast path (MX FP8 hardware) is available
*
* \tparam interp interpretation of fp8
* \param f 2xfloat
* \param scale scaling factor
* \return 2xfp8_storage_t
*/
template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
__host__ __device__ static inline fp8x2_storage_t cvt_float_to_fp8_scaled(const float2_t f,
float scale)
{
__is_interpret_supported(interp);
uint32_t rng = 0;
if constexpr(stochastic_rounding)
{
constexpr int seed = 1254739;
rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f[0]);
}
return cast_to_f8_from_f32_scaled<interp, stochastic_rounding>(f, rng, scale);
}
#else
/**
* \brief convert float to @p fp8_storage_t with scaling
*
* This version is used when the fast path (MX FP8 hardware) is not available
*
* \tparam interp interpretation of fp8
* \param f float number
* \param scale scaling factor
* \return fp8_storage_t
*/
template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
__host__ __device__ static inline fp8_storage_t cvt_float_to_fp8_scaled(const float f, float scale)
{
static_assert(interp == ck_fp8_interpretation_t::CK_E4M3_OCP ||
interp == ck_fp8_interpretation_t::CK_E5M2_OCP,
"Only OCP interpretations are supported");
uint32_t rng = 0;
if constexpr(stochastic_rounding)
{
constexpr int seed = 1254739;
rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f);
}
if constexpr(interp == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
return cast_to_f8<float, 3, 4, false, true, stochastic_rounding>(f / scale, rng);
}
else if constexpr(interp == ck_fp8_interpretation_t::CK_E5M2_OCP)
{
return cast_to_f8<float, 2, 5, false, true, stochastic_rounding>(f / scale, rng);
}
else
{
__hip_assert(false && "FP8 type is not supported by current target device");
return 0;
}
}
/**
* \brief convert two float to @p 2xfp8_storage_t with scaling
*
* This version is used when the fast path (MX FP8 hardware) is not available
*
* \tparam interp interpretation of fp8
* \param f 2xfloat
* \param scale scaling factor
* \return 2xfp8_storage_t
*/
template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
__host__ __device__ static inline fp8x2_storage_t cvt_float_to_fp8_scaled(const float2_t f,
float scale)
{
static_assert(interp == ck_fp8_interpretation_t::CK_E4M3_OCP ||
interp == ck_fp8_interpretation_t::CK_E5M2_OCP,
"Only OCP interpretations are supported");
uint32_t rng = 0;
if constexpr(stochastic_rounding)
{
constexpr int seed = 1254739;
rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f[0]);
}
if constexpr(interp == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
return {cast_to_f8<float, 3, 4, false, true, stochastic_rounding>(f[0] / scale, rng),
cast_to_f8<float, 3, 4, false, true, stochastic_rounding>(f[1] / scale, rng)};
}
else if constexpr(interp == ck_fp8_interpretation_t::CK_E5M2_OCP)
{
return {cast_to_f8<float, 2, 5, false, true, stochastic_rounding>(f[0] / scale, rng),
cast_to_f8<float, 2, 5, false, true, stochastic_rounding>(f[1] / scale, rng)};
}
else
{
__hip_assert(false && "FP8 type is not supported by current target device");
return 0;
}
}
#endif // CK_MX_FP8_CVT_FAST_PATH
} // namespace fp8_impl
// Declare a template function for fp8 conversion using SR
template <typename Y, typename X>
__host__ __device__ constexpr Y mxf8_convert_sr(X x, float scale);
// Declare a template function for fp8 conversion using RNE
template <typename Y, typename X>
__host__ __device__ constexpr Y mxf8_convert_rne(X x, float scale);
// convert fp32 to fp8 with rounding to nearest even
template <>
inline __host__ __device__ f8_ocp_t mxf8_convert_rne<f8_ocp_t, float>(float x, float scale)
{
return f8_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret>(x, scale)};
}
// convert fp32 to bf8 with rounding to nearest even
template <>
inline __host__ __device__ bf8_ocp_t mxf8_convert_rne<bf8_ocp_t, float>(float x, float scale)
{
return bf8_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret>(x, scale)};
}
// convert fp32x2 to fp8x2 with rounding to nearest even
template <>
inline __host__ __device__ f8x2_ocp_t mxf8_convert_rne<f8x2_ocp_t, float2_t>(float2_t x,
float scale)
{
return f8x2_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret>(x, scale)};
}
// convert fp32x2 to bf8x2 with rounding to nearest even
template <>
inline __host__ __device__ bf8x2_ocp_t mxf8_convert_rne<bf8x2_ocp_t, float2_t>(float2_t x,
float scale)
{
return bf8x2_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret>(x, scale)};
}
// convert fp32x16 to fp8x16 with rounding to nearest even
template <>
inline __host__ __device__ f8x16_ocp_t mxf8_convert_rne<f8x16_ocp_t, float16_t>(float16_t x,
float scale)
{
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} in{x};
union
{
f8x16_ocp_t fp8_1x16;
f8x2_ocp_t fp8_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}(
[&](auto i) { out.fp8_2x8[i] = mxf8_convert_rne<f8x2_ocp_t>(in.float_2x8[i], scale); });
return out.fp8_1x16;
}
// convert fp32x16 to bf8x16 with rounding to nearest even
template <>
inline __host__ __device__ bf8x16_ocp_t mxf8_convert_rne<bf8x16_ocp_t, float16_t>(float16_t x,
float scale)
{
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} in{x};
union
{
bf8x16_ocp_t bf8_1x16;
bf8x2_ocp_t bf8_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}(
[&](auto i) { out.bf8_2x8[i] = mxf8_convert_rne<bf8x2_ocp_t>(in.float_2x8[i], scale); });
return out.bf8_1x16;
}
// convert fp32x32 to fp8x32 with rounding to nearest even
template <>
inline __host__ __device__ f8x32_ocp_t mxf8_convert_rne<f8x32_ocp_t, float32_t>(float32_t x,
float scale)
{
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} in{x};
union
{
f8x32_ocp_t fp8_1x32;
f8x16_ocp_t fp8_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}(
[&](auto i) { out.fp8_16x2[i] = mxf8_convert_rne<f8x16_ocp_t>(in.float_16x2[i], scale); });
return out.fp8_1x32;
}
// convert fp32x32 to bf8x32 with rounding to nearest even
template <>
inline __host__ __device__ bf8x32_ocp_t mxf8_convert_rne<bf8x32_ocp_t, float32_t>(float32_t x,
float scale)
{
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} in{x};
union
{
bf8x32_ocp_t bf8_1x32;
bf8x16_ocp_t bf8_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}(
[&](auto i) { out.bf8_16x2[i] = mxf8_convert_rne<bf8x16_ocp_t>(in.float_16x2[i], scale); });
return out.bf8_1x32;
}
// convert fp32 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_ocp_t mxf8_convert_sr<f8_ocp_t, float>(float x, float scale)
{
return f8_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret, true>(x, scale)};
}
// convert fp32 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_ocp_t mxf8_convert_sr<bf8_ocp_t, float>(float x, float scale)
{
return bf8_ocp_t{
fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret, true>(x, scale)};
}
// convert fp32x2 to fp8x2 with stochastic rounding
template <>
inline __host__ __device__ f8x2_ocp_t mxf8_convert_sr<f8x2_ocp_t, float2_t>(float2_t x, float scale)
{
return f8x2_ocp_t{
fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret, true>(x, scale)};
}
// convert fp32x2 to bf8x2 with stochastic rounding
template <>
inline __host__ __device__ bf8x2_ocp_t mxf8_convert_sr<bf8x2_ocp_t, float2_t>(float2_t x,
float scale)
{
return bf8x2_ocp_t{
fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret, true>(x, scale)};
}
// convert fp32x16 to fp8x16 with stochastic rounding
template <>
inline __host__ __device__ f8x16_ocp_t mxf8_convert_sr<f8x16_ocp_t, float16_t>(float16_t x,
float scale)
{
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} in{x};
union
{
f8x16_ocp_t fp8_1x16;
f8x2_ocp_t fp8_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}(
[&](auto i) { out.fp8_2x8[i] = mxf8_convert_sr<f8x2_ocp_t>(in.float_2x8[i], scale); });
return out.fp8_1x16;
}
// convert fp32x16 to bf8x16 with stochastic rounding
template <>
inline __host__ __device__ bf8x16_ocp_t mxf8_convert_sr<bf8x16_ocp_t, float16_t>(float16_t x,
float scale)
{
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} in{x};
union
{
bf8x16_ocp_t bf8_1x16;
bf8x2_ocp_t bf8_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}(
[&](auto i) { out.bf8_2x8[i] = mxf8_convert_sr<bf8x2_ocp_t>(in.float_2x8[i], scale); });
return out.bf8_1x16;
}
// convert fp32x32 to fp8x32 with stochastic rounding
template <>
inline __host__ __device__ f8x32_ocp_t mxf8_convert_sr<f8x32_ocp_t, float32_t>(float32_t x,
float scale)
{
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} in{x};
union
{
f8x32_ocp_t fp8_1x32;
f8x16_ocp_t fp8_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}(
[&](auto i) { out.fp8_16x2[i] = mxf8_convert_sr<f8x16_ocp_t>(in.float_16x2[i], scale); });
return out.fp8_1x32;
}
// convert fp32x32 to bf8x32 with stochastic rounding
template <>
inline __host__ __device__ bf8x32_ocp_t mxf8_convert_sr<bf8x32_ocp_t, float32_t>(float32_t x,
float scale)
{
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} in{x};
union
{
bf8x32_ocp_t bf8_1x32;
bf8x16_ocp_t bf8_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}(
[&](auto i) { out.bf8_16x2[i] = mxf8_convert_sr<bf8x16_ocp_t>(in.float_16x2[i], scale); });
return out.bf8_1x32;
}
} // namespace ck
include/ck/utility/mxfp_utils.hpp 0 → 100644
View file @ 9ba504b6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck::utils {
union cvt
{
float value_float;
uint32_t value_bitwise;
};
template <typename DTYPE>
inline bool getDataHasInf()
{
return DTYPE::dataInfo.hasInf;
}
template <typename T>
__host__ __device__ inline bool is_zero(e8m0_bexp_t const scale, T const data);
template <typename T>
__host__ __device__ inline bool is_nan(e8m0_bexp_t const scale, T const data);
template <typename T>
__host__ __device__ inline bool is_inf(e8m0_bexp_t const scale, T const data);
template <typename T>
__host__ __device__ inline int get_exponent_value(T x)
{
x >>= NumericUtils<T>::mant;
x &= ((1 << NumericUtils<T>::exp) - 1);
return static_cast<int>(x);
}
template <typename T>
__host__ __device__ inline bool is_subnormal(T x)
{
return get_exponent_value<T>(x) == 0;
}
template <typename T>
__host__ __device__ inline double get_mantissa_value(T x)
{
double mantissa = is_subnormal<T>(x) ? 0.0f : 1.0f;
for(uint i = 0; i < NumericUtils<T>::mant; i++)
{
mantissa += std::pow(2, -int32_t((NumericUtils<T>::mant - i))) * (x & 0b1);
x >>= 1;
}
return mantissa;
}
template <typename T>
__host__ __device__ inline bool get_data_has_inf()
{
return NumericUtils<T>::has_inf;
}
template <typename T>
__host__ __device__ float convert_to_float(T data, int scale_exp)
{
float d_sign =
std::pow(-1, static_cast<float>(data >> (NumericUtils<T>::exp + NumericUtils<T>::mant)));
float d_exp;
if(is_subnormal<T>(data))
d_exp = std::pow(2, 1 - static_cast<int>(NumericUtils<T>::bias));
else
d_exp = std::pow(2, get_exponent_value<T>(data) - static_cast<int>(NumericUtils<T>::bias));
float d_mant = get_mantissa_value<T>(data);
float data_value = d_sign * d_exp * d_mant;
float scale_value = std::pow(
2, static_cast<float>((scale_exp - static_cast<int>(NumericUtils<e8m0_bexp_t>::bias))));
return data_value * scale_value;
}
template <typename T>
__host__ __device__ inline float to_float(e8m0_bexp_t const scale, T const data);
template <typename T>
__host__ __device__ T sat_convert_to_type(float value);
template <typename T>
__host__ __device__ T sat_convert_to_type_sr(float value, uint32_t seed);
template <typename T>
inline T convert_to_type(float value)
{
using bitwise_type = typename NumericUtils<T>::bitwise_type;
if(std::abs(value) > NumericLimits<T>::Max())
{
float max_value = NumericLimits<T>::Max();
cvt t;
// cppcheck-suppress redundantAssignment
t.value_float = max_value;
uint32_t max_bitwise = t.value_bitwise;
// cppcheck-suppress redundantAssignment
t.value_float = value;
bitwise_type sign =
t.value_bitwise >> (NumericUtils<float>::exp + NumericUtils<float>::mant);
bitwise_type exp =
((max_bitwise >> NumericUtils<float>::mant) & NumericUtils<float>::exp_mask) -
(NumericUtils<float>::bias - NumericUtils<T>::bias);
bitwise_type mantissa = max_bitwise >> (NumericUtils<float>::mant - NumericUtils<T>::mant);
uint32_t mant_prev = max_bitwise >> (NumericUtils<float>::mant - NumericUtils<T>::mant);
mant_prev &= ((1 << NumericUtils<T>::mant) - 1);
mant_prev--;
mant_prev <<= (NumericUtils<float>::mant - NumericUtils<T>::mant);
uint32_t prev_bit =
((max_bitwise >> NumericUtils<float>::mant) << NumericUtils<float>::mant) | mant_prev;
t.value_bitwise = prev_bit;
float prev_val = t.value_float;
float diff = max_value - prev_val;
float actual_max = max_value + (diff / 2);
if(std::abs(value) < actual_max)
{
return sign << ((NumericUtils<T>::exp + NumericUtils<T>::mant)) |
(exp << NumericUtils<T>::mant) | mantissa;
}
else
{
if(!get_data_has_inf<T>())
{
return (1 << (NumericUtils<T>::mant + NumericUtils<T>::exp)) - 1;
}
else
{
exp++;
return sign << ((NumericUtils<T>::exp + NumericUtils<T>::mant)) |
(exp << NumericUtils<T>::mant);
}
}
}
const int mfmt = NumericUtils<float>::mant;
uint32_t x;
x = bit_cast<uint32_t>(value);
uint32_t head, mantissa;
int32_t exponent, bias;
uint32_t sign;
head = x & NumericUtils<float>::head_mask;
mantissa = x & NumericUtils<float>::mant_mask;
exponent = (head >> NumericUtils<float>::mant) & NumericUtils<float>::exp_mask;
sign = head >> (NumericUtils<float>::mant + NumericUtils<float>::exp);
bias = NumericUtils<float>::bias;
if(x == 0)
{
return 0b0;
}
const int mini_bias = NumericUtils<T>::bias;
const int mini_denormal_act_exponent = 1 - mini_bias;
int act_exponent, out_exponent, exponent_diff;
bool is_subnorm = false;
if(exponent == 0)
{
act_exponent = exponent - bias + 1;
exponent_diff = mini_denormal_act_exponent - act_exponent;
is_subnorm = true;
}
else
{
act_exponent = exponent - bias;
if(act_exponent <= mini_denormal_act_exponent)
{
exponent_diff = mini_denormal_act_exponent - act_exponent;
is_subnorm = true;
}
else
{
exponent_diff = 0;
}
mantissa += (1UL << mfmt);
}
auto shift_amount = (mfmt - NumericUtils<T>::mant + exponent_diff);
shift_amount = (shift_amount >= 64) ? 63 : shift_amount;
bool midpoint = (mantissa & ((1UL << shift_amount) - 1)) == (1UL << (shift_amount - 1));
float min_subnorm = NumericLimits<T>::DataMinSubnorm() * (sign ? -1 : 1);
if(is_subnorm && std::abs(value) < std::abs(min_subnorm))
{
// closer to 0
if(std::abs(value) <= std::abs(min_subnorm - value))
return 0;
else
return 1 | (sign << (NumericUtils<T>::exp + NumericUtils<T>::mant));
}
if(exponent_diff > 0)
mantissa >>= exponent_diff;
else if(exponent_diff == -1)
mantissa <<= -exponent_diff;
bool implicit_one = mantissa & (1 << mfmt);
out_exponent = (act_exponent + exponent_diff) + mini_bias - (implicit_one ? 0 : 1);
uint32_t drop_mask = (1UL << (mfmt - NumericUtils<T>::mant)) - 1;
bool odd = mantissa & (1UL << (mfmt - NumericUtils<T>::mant));
mantissa += (midpoint ? (odd ? mantissa : mantissa - 1) : mantissa) & drop_mask;
if(out_exponent == 0)
{
if((1UL << mfmt) & mantissa)
{
out_exponent = 1;
}
}
else
{
if((1UL << (mfmt + 1)) & mantissa)
{
mantissa >>= 1;
out_exponent++;
}
}
mantissa >>= (mfmt - NumericUtils<T>::mant);
if(out_exponent == 0 && mantissa == 0)
{
return 0;
}
mantissa &= (1UL << NumericUtils<T>::mant) - 1;
return (sign << (NumericUtils<T>::exp + NumericUtils<T>::mant)) |
(out_exponent << NumericUtils<T>::mant) | mantissa;
}
template <typename T>
inline T convert_to_type_sr(float value, uint32_t seed)
{
if(std::abs(value) > NumericLimits<T>::Max())
{
float max_value = NumericLimits<T>::Max();
cvt t;
// cppcheck-suppress redundantAssignment
t.value_float = max_value;
uint max_bitwise = t.value_bitwise;
// cppcheck-suppress redundantAssignment
t.value_float = value;
T sign = t.value_bitwise >> (NumericUtils<float>::exp + NumericUtils<float>::mant);
T exp = ((max_bitwise >> NumericUtils<float>::mant) & NumericUtils<float>::exp_mask) -
(NumericUtils<float>::bias - NumericUtils<T>::bias);
uint32_t mant_prev = max_bitwise >> (NumericUtils<float>::mant - NumericUtils<T>::mant);
mant_prev &= ((1UL << NumericUtils<T>::mant) - 1);
mant_prev--;
mant_prev <<= (NumericUtils<float>::mant - NumericUtils<T>::mant);
uint32_t prev_bit =
((max_bitwise >> NumericUtils<float>::mant) << NumericUtils<float>::mant) | mant_prev;
t.value_bitwise = prev_bit;
float prev_val = t.value_float;
float diff = max_value - prev_val;
float actual_max = max_value + (diff / 2);
if(std::abs(value) < actual_max)
{
double d_max_value = static_cast<double>(max_value);
double d_actual_max = static_cast<double>(actual_max);
double d_value = static_cast<double>(value);
double d_is = std::abs(d_max_value - d_actual_max);
double d_seed = static_cast<double>(seed);
double d_prob = 1.0f - (std::abs(d_value - d_max_value) / d_is); // prob to round down
double thresh = UINT_MAX * d_prob;
if(!get_data_has_inf<T>() || d_seed <= thresh)
// return static_cast<T>(satConvertToType(getDataMax<DTYPE>())); //round down time
return sign == 0 ? NumericUtils<f4_t>::data_max_positive_normal_mask
: NumericUtils<f4_t>::data_max_negative_normal_mask;
else
{
exp++;
return sign << ((NumericUtils<T>::exp + NumericUtils<T>::mant)) // inf
| (exp << NumericUtils<T>::mant);
}
}
else
{
if(!get_data_has_inf<T>())
return (1 << (NumericUtils<T>::mant + NumericUtils<T>::exp)) - 1;
else
{
exp++;
return sign << ((NumericUtils<T>::exp + NumericUtils<T>::mant)) // inf
| (exp << NumericUtils<T>::mant);
}
}
}
uint32_t f32 = bit_cast<uint32_t>(value);
auto f32_mant = f32 & NumericUtils<float>::mant_mask;
auto head = f32 & NumericUtils<float>::head_mask;
auto f32_exp = (head >> NumericUtils<float>::mant) & NumericUtils<float>::exp_mask;
auto sign_bit = head >> (NumericUtils<float>::mant + NumericUtils<float>::exp);
auto sign = sign_bit << (NumericUtils<T>::exp + NumericUtils<T>::mant);
f32_exp = static_cast<int32_t>(f32_exp) - NumericUtils<float>::bias;
int32_t exp = f32_exp;
auto mant = f32_mant;
bool subnorm = false;
if(f32 == 0)
return 0b0;
if(exp >= NumericUtils<T>::unbiased_exp_min)
{
mant = f32_mant;
}
// if the exponent bit is 8, then the subnormal is exactly the same as f32
else if(exp < NumericUtils<T>::unbiased_exp_min &&
NumericUtils<T>::exp < NumericUtils<float>::exp)
{
subnorm = true;
auto diff = static_cast<uint32_t>(NumericUtils<T>::unbiased_exp_min - exp);
if(diff >= 32)
{
mant = 0;
f32_mant = 0;
}
else
{
f32_mant |= static_cast<uint32_t>(1) << NumericUtils<float>::mant;
f32_mant >>= diff;
}
exp = 0;
mant = f32_mant;
}
uint32_t sr_shift = NumericUtils<T>::sr_shift;
// For stochastic-rounding we add the aligned random value to the
// mantissa and then truncate (RTZ).
mant += seed >> sr_shift;
// Increment exponent when mantissa overflows due to rounding
if(mant >= static_cast<uint32_t>(1) << NumericUtils<float>::mant)
++exp;
mant >>= (NumericUtils<float>::mant - NumericUtils<T>::mant);
mant &= ((1 << NumericUtils<T>::mant) - 1);
auto biased_exp = static_cast<uint32_t>(exp);
if(!subnorm)
biased_exp = static_cast<uint32_t>(exp + NumericUtils<T>::bias);
biased_exp &= ((1 << NumericUtils<T>::exp) - 1);
auto val = sign | biased_exp << NumericUtils<T>::mant | mant;
return val;
}
} // namespace ck::utils
include/ck/utility/scaled_type_convert.hpp 0 → 100644
View file @ 9ba504b6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/type_convert.hpp"
#include "ck/utility/mxf8_utils.hpp"
#ifdef CK_USE_NATIVE_MX_SUPPORT
#define CK_USE_NATIVE_MX_SUPPORT 1
#else
#define CK_USE_NATIVE_MX_SUPPORT 0
#endif
namespace ck {
// Declare a template function for scaled conversion
template <typename Y, typename X>
#if CK_USE_OCP_FP8
__host__ __device__ constexpr Y scaled_type_convert(e8m0_bexp_t scale, X x);
#else
__host__ constexpr Y scaled_type_convert(e8m0_bexp_t scale, X x);
#endif
// convert f8_ocp_t to fp32
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float scaled_type_convert<float, f8_ocp_t>(e8m0_bexp_t scale, f8_ocp_t x)
#else
inline __host__ float scaled_type_convert<float, f8_ocp_t>(e8m0_bexp_t scale, f8_ocp_t x)
#endif
{
#if CK_MX_FP8_CVT_FAST_PATH
return fp8_impl::cast_to_f32_from_f8_scaled<f8_ocp_t::default_interpret>(
type_convert<float>(scale), x.data);
#else
return type_convert<float>(scale) * type_convert<float>(x);
#endif
}
// convert bf8_ocp_t to fp32
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float scaled_type_convert<float, bf8_ocp_t>(e8m0_bexp_t scale,
bf8_ocp_t x)
#else
inline __host__ float scaled_type_convert<float, bf8_ocp_t>(e8m0_bexp_t scale, bf8_ocp_t x)
#endif
{
#if CK_MX_FP8_CVT_FAST_PATH
return fp8_impl::cast_to_f32_from_f8_scaled<bf8_ocp_t::default_interpret>(
type_convert<float>(scale), x.data);
#else
return type_convert<float>(scale) * type_convert<float>(x);
#endif
}
// convert 2 x f8_ocp_t to 2 x fp32
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float2_t scaled_type_convert<float2_t, f8x2_ocp_t>(e8m0_bexp_t scale,
f8x2_ocp_t x)
#else
inline __host__ float2_t scaled_type_convert<float2_t, f8x2_ocp_t>(e8m0_bexp_t scale, f8x2_ocp_t x)
#endif
{
#if CK_MX_FP8_CVT_FAST_PATH
return fp8_impl::cast_to_f32x2_from_f8x2_scaled<f8_ocp_t::default_interpret>(
type_convert<float>(scale), x.AsType<fp8_impl::fp8x2_storage_t>()[Number<0>{}]);
#else
return float2_t{scaled_type_convert<float>(scale, x.AsType<f8_ocp_t>()[Number<0>{}]),
scaled_type_convert<float>(scale, x.AsType<f8_ocp_t>()[Number<1>{}])};
#endif
}
// convert 2 x bf8_ocp_t to 2 x fp32
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float2_t scaled_type_convert<float2_t, bf8x2_ocp_t>(e8m0_bexp_t scale,
bf8x2_ocp_t x)
#else
inline __host__ float2_t scaled_type_convert<float2_t, bf8x2_ocp_t>(e8m0_bexp_t scale,
bf8x2_ocp_t x)
#endif
{
#if CK_MX_FP8_CVT_FAST_PATH
return fp8_impl::cast_to_f32x2_from_f8x2_scaled<bf8_ocp_t::default_interpret>(
type_convert<float>(scale), x.AsType<fp8_impl::fp8x2_storage_t>()[Number<0>{}]);
#else
return float2_t{scaled_type_convert<float>(scale, x.AsType<bf8_ocp_t>()[Number<0>{}]),
scaled_type_convert<float>(scale, x.AsType<bf8_ocp_t>()[Number<1>{}])};
#endif
}
// convert 16 x f8_ocp_t to 16 x fp32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float16_t scaled_type_convert<float16_t, f8x16_ocp_t>(e8m0_bexp_t scale,
f8x16_ocp_t x)
#else
inline __host__ float16_t scaled_type_convert<float16_t, f8x16_ocp_t>(e8m0_bexp_t scale,
f8x16_ocp_t x)
#endif
{
union
{
f8x16_ocp_t f8_1x16;
f8x2_ocp_t f8_2x8[8];
} in{x};
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}([&](auto i) {
out.float_2x8[i] = scaled_type_convert<float2_t, f8x2_ocp_t>(scale, in.f8_2x8[i]);
});
return out.float_1x16;
}
// convert 16 x bf8_ocp_t to 16 x fp32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float16_t scaled_type_convert<float16_t, bf8x16_ocp_t>(e8m0_bexp_t scale,
bf8x16_ocp_t x)
#else
inline __host__ float16_t scaled_type_convert<float16_t, bf8x16_ocp_t>(e8m0_bexp_t scale,
bf8x16_ocp_t x)
#endif
{
union
{
bf8x16_ocp_t bf8_1x16;
bf8x2_ocp_t bf8_2x8[8];
} in{x};
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}([&](auto i) {
out.float_2x8[i] = scaled_type_convert<float2_t, bf8x2_ocp_t>(scale, in.bf8_2x8[i]);
});
return out.float_1x16;
}
// convert 32 x f8_ocp_t to 32 x fp32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float32_t scaled_type_convert<float32_t, f8x32_ocp_t>(e8m0_bexp_t scale,
f8x32_ocp_t x)
#else
inline __host__ float32_t scaled_type_convert<float32_t, f8x32_ocp_t>(e8m0_bexp_t scale,
f8x32_ocp_t x)
#endif
{
union
{
f8x32_ocp_t f8_1x32;
f8x16_ocp_t f8_16x2[2];
} in{x};
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}([&](auto i) {
out.float_16x2[i] = scaled_type_convert<float16_t, f8x16_ocp_t>(scale, in.f8_16x2[i]);
});
return out.float_1x32;
}
// convert 32 x bf8_ocp_t to 32 x fp32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float32_t scaled_type_convert<float32_t, bf8x32_ocp_t>(e8m0_bexp_t scale,
bf8x32_ocp_t x)
#else
inline __host__ float32_t scaled_type_convert<float32_t, bf8x32_ocp_t>(e8m0_bexp_t scale,
bf8x32_ocp_t x)
#endif
{
union
{
bf8x32_ocp_t bf8_1x32;
bf8x16_ocp_t bf8_16x2[2];
} in{x};
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}([&](auto i) {
out.float_16x2[i] = scaled_type_convert<float16_t, bf8x16_ocp_t>(scale, in.bf8_16x2[i]);
});
return out.float_1x32;
}
// convert fp32 to fp8
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ f8_ocp_t scaled_type_convert<f8_ocp_t, float>(e8m0_bexp_t scale, float x)
#else
inline __host__ f8_ocp_t scaled_type_convert<f8_ocp_t, float>(e8m0_bexp_t scale, float x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<f8_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<f8_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32 to bf8
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ bf8_ocp_t scaled_type_convert<bf8_ocp_t, float>(e8m0_bexp_t scale,
float x)
#else
inline __host__ bf8_ocp_t scaled_type_convert<bf8_ocp_t, float>(e8m0_bexp_t scale, float x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<bf8_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<bf8_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x2 to fp8x2
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ f8x2_ocp_t scaled_type_convert<f8x2_ocp_t, float2_t>(e8m0_bexp_t scale,
float2_t x)
#else
inline __host__ f8x2_ocp_t scaled_type_convert<f8x2_ocp_t, float2_t>(e8m0_bexp_t scale, float2_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<f8x2_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<f8x2_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x2 to bf8x2
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ bf8x2_ocp_t scaled_type_convert<bf8x2_ocp_t, float2_t>(e8m0_bexp_t scale,
float2_t x)
#else
inline __host__ bf8x2_ocp_t scaled_type_convert<bf8x2_ocp_t, float2_t>(e8m0_bexp_t scale,
float2_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<bf8x2_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<bf8x2_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x16 to fp8x16
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ f8x16_ocp_t
scaled_type_convert<f8x16_ocp_t, float16_t>(e8m0_bexp_t scale, float16_t x)
#else
inline __host__ f8x16_ocp_t scaled_type_convert<f8x16_ocp_t, float16_t>(e8m0_bexp_t scale,
float16_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<f8x16_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<f8x16_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x16 to bf8x16
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ bf8x16_ocp_t
scaled_type_convert<bf8x16_ocp_t, float16_t>(e8m0_bexp_t scale, float16_t x)
#else
inline __host__ bf8x16_ocp_t scaled_type_convert<bf8x16_ocp_t, float16_t>(e8m0_bexp_t scale,
float16_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<bf8x16_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<bf8x16_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x32 to fp8x32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ f8x32_ocp_t
scaled_type_convert<f8x32_ocp_t, float32_t>(e8m0_bexp_t scale, float32_t x)
#else
inline __host__ f8x32_ocp_t scaled_type_convert<f8x32_ocp_t, float32_t>(e8m0_bexp_t scale,
float32_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<f8x32_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<f8x32_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x32 to bf8x32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ bf8x32_ocp_t
scaled_type_convert<bf8x32_ocp_t, float32_t>(e8m0_bexp_t scale, float32_t x)
#else
inline __host__ bf8x32_ocp_t scaled_type_convert<bf8x32_ocp_t, float32_t>(e8m0_bexp_t scale,
float32_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<bf8x32_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<bf8x32_ocp_t>(x, type_convert<float>(scale));
#endif
}
// activate for architectures with native MX support
#if CK_USE_NATIVE_MX_SUPPORT
// convert fp4 to fp32
template <>
inline __host__ __device__ float scaled_type_convert<float, f4_t>(e8m0_bexp_t scale, f4_t x)
{
#if defined(__gfx950__)
union
{
float float_array[2];
float2_t float2_array;
} float_values{};
float_values.float2_array =
__builtin_amdgcn_cvt_scalef32_pk_f32_fp4(x, type_convert<float>(scale), 0);
return float_values.float_array[0];
#else
return utils::to_float<f4_t>(scale, x);
#endif
}
// convert vector of 2 fp4 to vector of 2 fp32
template <>
inline __host__ __device__ float2_t scaled_type_convert<float2_t, f4x2_t>(e8m0_bexp_t scale,
f4x2_t x)
{
#if defined(__gfx950__)
union
{
uint32_t bitwise;
f4x2_t f4x2_array[4];
} value{};
value.f4x2_array[0] = x;
return __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(value.bitwise, type_convert<float>(scale), 0);
#else
float2_t ret{utils::to_float<f4_t>(
scale, x.template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{})),
utils::to_float<f4_t>(
scale, x.template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}))};
return ret;
#endif
}
// convert vector of 32 fp4 to vector of 32 fp32
template <>
inline __host__ __device__ float32_t scaled_type_convert<float32_t, f4x32_t>(e8m0_bexp_t scale,
f4x32_t x)
{
#if defined(__gfx950__)
union
{
f4x32_t f4x32_array;
f4x2_t fp4x2[16];
} value{x};
union
{
uint32_t bitwise;
f4x2_t f4x2_array[4];
} bitwise_value{};
float2_t op;
float32_t ret;
// TODO: pack in a loop
bitwise_value.f4x2_array[0] = value.fp4x2[0];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[0] = op[0];
ret[1] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[1];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[2] = op[0];
ret[3] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[2];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[4] = op[0];
ret[5] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[3];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[6] = op[0];
ret[7] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[4];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[8] = op[0];
ret[9] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[5];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[10] = op[0];
ret[11] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[6];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[12] = op[0];
ret[13] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[7];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[14] = op[0];
ret[15] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[8];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[16] = op[0];
ret[17] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[9];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[18] = op[0];
ret[19] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[10];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[20] = op[0];
ret[21] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[11];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[22] = op[0];
ret[23] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[12];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[24] = op[0];
ret[25] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[13];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[26] = op[0];
ret[27] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[14];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[28] = op[0];
ret[29] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[15];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[30] = op[0];
ret[31] = op[1];
return ret;
#else
union
{
float32_t float32_array;
float float_array[32];
} float_values{};
union
{
__uint128_t bitwise;
f4x2_t f4x2_array[16];
f4x32_t f4x32_array;
} f4_values{bit_cast<__uint128_t>(x)};
// TODO: pack in a loop
float_values.float_array[0] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[0].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[0].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[1].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[1].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[2].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[2].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[3].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[3].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[0] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[4].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[4].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[5].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[5].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[6].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[6].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[7].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[7].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[0] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[8].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[8].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[9].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[9].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[10].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[10].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[11].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[11].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[0] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[12].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[12].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[13].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[13].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[14].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[14].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[15].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[15].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
return float_values.float32_array;
#endif
}
// convert fp32 to fp4
template <>
inline __host__ __device__ f4_t scaled_type_convert<f4_t, float>(e8m0_bexp_t scale, float x)
{
#if CK_USE_SR_F4_CONVERSION
return f4_convert_sr(x, type_convert<float>(scale));
#else
return f4_convert_rne(x, type_convert<float>(scale));
#endif
}
// convert vector of 2 fp32 to vector of 2 fp4
template <>
inline __host__ __device__ f4x2_t scaled_type_convert<f4x2_t, float2_t>(e8m0_bexp_t scale,
float2_t x)
{
#if CK_USE_SR_F4_CONVERSION
return f4_convert_sr(x, type_convert<float>(scale));
#else
return f4_convert_rne(x, type_convert<float>(scale));
#endif
}
// convert vector of 32 fp32 to vector of 32 fp4
template <>
inline __host__ __device__ f4x32_t scaled_type_convert<f4x32_t, float32_t>(e8m0_bexp_t scale,
float32_t x)
{
#if CK_USE_SR_F4_CONVERSION
return f4_convert_sr(x, type_convert<float>(scale));
#else
return f4_convert_rne(x, type_convert<float>(scale));
#endif
}
/**
* @brief Converts a 6-bit floating-point value (f6_t) to a 32-bit float,
* applying the specified scaling factor.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param x The f6_t value to be converted.
* @return The converted 32-bit float representation of the input.
*/
template <>
inline __host__ __device__ float scaled_type_convert<float, f6_t>(e8m0_bexp_t scale, f6_t x)
{
#if defined(__gfx950__)
union
{
f6x32_t f6_vector;
f6_t f6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
out.float_vector =
__builtin_amdgcn_cvt_scalef32_pk32_f32_fp6(in.f6_vector, type_convert<float>(scale));
return out.float_array[0];
#else
return utils::to_float<f6_t>(scale, x);
#endif
}
/**
* @brief Converts a vector of 32 6-bit floating-point values (f6x32_t) to a vector of 32 floats,
* applying the specified scaling factor.
*
* @param scale The exponent scale factor (e8m0_bexp_t).
* @param x The f6x32_t vector to be converted.
* @return The converted float vector representation of the input.
*/
template <>
inline __host__ __device__ float32_t scaled_type_convert<float32_t, f6x32_t>(e8m0_bexp_t scale,
f6x32_t x)
{
#if defined(__gfx950__)
return __builtin_amdgcn_cvt_scalef32_pk32_f32_fp6(x, type_convert<float>(scale));
#else
union
{
f6x32_t f6_vector;
f6_t f6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
ck::static_for<0, 32, 1>{}(
[&](auto i) { out.float_array[i] = utils::to_float<f6_t>(scale, in.f6_array[i]); });
return out.float_vector;
#endif
}
/**
* @brief Converts a 6-bit floating-point value (bf6_t) to a 32-bit float,
* applying the specified scaling factor.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param x The bf6_t value to be converted.
* @return The converted 32-bit float representation of the input.
*/
template <>
inline __host__ __device__ float scaled_type_convert<float, bf6_t>(e8m0_bexp_t scale, bf6_t x)
{
#if defined(__gfx950__)
union
{
bf6x32_t bf6_vector;
bf6_t bf6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
out.float_vector =
__builtin_amdgcn_cvt_scalef32_pk32_f32_bf6(in.bf6_vector, type_convert<float>(scale));
return out.float_array[0];
#else
return utils::to_float<bf6_t>(scale, x);
#endif
}
/**
* @brief Converts a vector of 6-bit floating-point values (bf6x32_t) to a vector of 32 floats,
* applying the specified scaling factor.
*
* @param scale The exponent scale factor (e8m0_bexp_t).
* @param x The bf6x32_t vector to be converted.
* @return The converted vector of 32 float representation of the input.
*/
template <>
inline __host__ __device__ float32_t scaled_type_convert<float32_t, bf6x32_t>(e8m0_bexp_t scale,
bf6x32_t x)
{
#if defined(__gfx950__)
return __builtin_amdgcn_cvt_scalef32_pk32_f32_bf6(x, type_convert<float>(scale));
#else
union
{
bf6x32_t bf6_vector;
bf6_t bf6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
ck::static_for<0, 32, 1>{}(
[&](auto i) { out.float_array[i] = utils::to_float<bf6_t>(scale, in.bf6_array[i]); });
return out.float_vector;
#endif
}
/**
* @brief Converts a 32-bit float to a 6-bit floating-point value (f6_t), applying the specified
* scale.
*
* Depending on whether CK_USE_SR_F6_CONVERSION is defined, it uses either stochastic rounding
* (f6_convert_sr) or round-to-nearest-even (f6_convert_rne).
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param x The float value to convert.
* @return The converted 6-bit floating-point value (f6_t).
*/
template <>
inline __host__ __device__ f6_t scaled_type_convert<f6_t, float>(e8m0_bexp_t scale, float x)
{
#if CK_USE_SR_F6_CONVERSION
return f6_convert_sr(x, type_convert<float>(scale));
#else
return f6_convert_rne(x, type_convert<float>(scale));
#endif
}
/**
* @brief Converts a vector of 32 floats to a vector of 32 6-bit floating-point values (f6x32_t),
* applying the specified scale.
*
* Depending on whether CK_USE_SR_F6_CONVERSION is defined, it uses either stochastic rounding
* (f6_convert_sr) or round-to-nearest-even (f6_convert_rne).
*
* @param scale The exponent scale factor (e8m0_bexp_t).
* @param x The float vector to convert.
* @return The converted vector of 6-bit floating-point values (f6x32_t).
*/
template <>
inline __host__ __device__ f6x32_t scaled_type_convert<f6x32_t, float32_t>(e8m0_bexp_t scale,
float32_t x)
{
#if CK_USE_SR_F6_CONVERSION
return f6_convert_sr(x, type_convert<float>(scale));
#else
return f6_convert_rne(x, type_convert<float>(scale));
#endif
}
/**
* @brief Converts a 32-bit float to a 6-bit floating-point value (bf6_t), applying the specified
* scale.
*
* Depending on whether CK_USE_SR_F6_CONVERSION is defined, it uses either stochastic rounding
* (bf6_convert_sr) or round-to-nearest-even (bf6_convert_rne).
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param x The float value to convert.
* @return The converted 6-bit floating-point value (bf6_t).
*/
template <>
inline __host__ __device__ bf6_t scaled_type_convert<bf6_t, float>(e8m0_bexp_t scale, float x)
{
#if CK_USE_SR_F6_CONVERSION
return bf6_convert_sr(x, type_convert<float>(scale));
#else
return bf6_convert_rne(x, type_convert<float>(scale));
#endif
}
/**
* @brief Converts a vector of 32 floats to a vector of 32 6-bit floating-point values (bf6x32_t),
* applying the specified scale.
*
* Depending on whether CK_USE_SR_F6_CONVERSION is defined, it uses either stochastic rounding
* (bf6_convert_sr) or round-to-nearest-even (bf6_convert_rne).
*
* @param scale The exponent scale factor (e8m0_bexp_t).
* @param x The float vector to convert.
* @return The converted 6-bit floating-point vector (bf6x32_t).
*/
template <>
inline __host__ __device__ bf6x32_t scaled_type_convert<bf6x32_t, float32_t>(e8m0_bexp_t scale,
float32_t x)
{
#if CK_USE_SR_F6_CONVERSION
return bf6_convert_sr(x, type_convert<float>(scale));
#else
return bf6_convert_rne(x, type_convert<float>(scale));
#endif
}
#endif // #if CK_USE_NATIVE_MX_SUPPORT
} // namespace ck
include/ck/utility/type_convert.hpp
View file @ 9ba504b6
......@@ -5,15 +5,39 @@
#include "ck/utility/data_type.hpp"
#include "ck/utility/f8_utils.hpp"
#include "ck/utility/mxf4_utils.hpp"
#include "ck/utility/mxf6_utils.hpp"
#include "ck/utility/random_gen.hpp"
#include "ck/utility/array.hpp"
#include "ck/utility/amd_inline_asm.hpp"
#include "ck/utility/type.hpp"
namespace ck {
// Define the common macro for MI300 models
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
#define __gfx94__
#endif
namespace {
namespace details {
[[maybe_unused]] __host__ half2_t pk_add_f16(const half2_t& x, const half2_t& y)
{
half2_t vector_res;
vector_res.x = x.x + y.x;
vector_res.y = x.y + y.y;
return vector_res;
}
[[maybe_unused]] __device__ half2_t pk_add_f16(const half2_t& x, const half2_t& y)
{
return amd_assembly_pk_add_f16(x, y);
}
} // namespace details
} // namespace
// Declare a template function for bf16 conversion using RTN
template <typename Y, typename X>
__host__ __device__ constexpr Y bf16_convert_rtn(X x);
......@@ -520,13 +544,51 @@ template <>
inline __host__ __device__ float2_t type_convert<float2_t, pk_i4_t>(pk_i4_t x)
{
uint8_t x_u8 = ck::bit_cast<uint8_t>(x);
uint8_t x_l = (x_u8 & 0x0f) >> 0;
uint8_t x_h = (x_u8 & 0xf0) >> 4;
auto l_f32 = ck::type_convert<float>(x_l);
auto h_f32 = ck::type_convert<float>(x_h);
float x_l = ((x_u8 & 0x0f) >> 0) - 8.f;
float x_h = ((x_u8 & 0xf0) >> 4) - 8.f;
#ifdef CK_USE_PK4_LAYOUT_SHUFFLE
float2_t res = {x_h, x_l};
#elif
float2_t res = {x_l, x_h};
#endif
return res;
}
template <>
inline __host__ __device__ half2_t type_convert<half2_t, pk_i4_t>(pk_i4_t x)
{
uint8_t x_u8 = ck::bit_cast<uint8_t>(x);
#ifdef CK_USE_PK4_LAYOUT_SHUFFLE
uint32_t i4s = ((x_u8 & 0x0f) << 16) | ((x_u8 & 0xf0) >> 4);
#else
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 {l_f32, h_f32};
return details::pk_add_f16(bit_cast<half2_t>(lo), bit_cast<half2_t>(SUB));
}
template <>
inline __host__ __device__ bhalf2_t type_convert<bhalf2_t, pk_i4_t>(pk_i4_t x)
{
uint8_t x_u8 = ck::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_USE_PK4_LAYOUT_SHUFFLE
bhalf2_t res = {type_convert<bhalf_t>(x_h), type_convert<bhalf_t>(x_l)};
#else
bhalf2_t res = {type_convert<bhalf_t>(x_l), type_convert<bhalf_t>(x_h)};
#endif
return res;
}
template <>
......@@ -647,6 +709,1278 @@ inline __host__ __device__ half_t type_convert<half_t, bf8_fnuz_t>(bf8_fnuz_t x)
#endif
}
// convert fp32 to fp4 with rounding to nearest even
inline __host__ __device__ f4_t f4_convert_rne(float x, float scale = 1.0f)
{
#if defined(__gfx950__)
union
{
uint32_t bitwise;
f4_t f4_array[4];
} value{0};
value.bitwise = __builtin_amdgcn_cvt_scalef32_pk_fp4_f32(value.bitwise, x, x, scale, 0);
return value.f4_array[0];
#else
return utils::sat_convert_to_type<f4_t>(x / scale);
#endif
}
// convert vector of 2 fp32 to vector of 2 fp4 with rne
inline __host__ __device__ f4x2_t f4_convert_rne(float2_t x, float scale = 1.0f)
{
#if defined(__gfx950__)
union
{
uint32_t bitwise;
f4x2_t f4x2_array[4];
} value{0};
value.bitwise = __builtin_amdgcn_cvt_scalef32_pk_fp4_f32(value.bitwise, x[0], x[1], scale, 0);
return value.f4x2_array[0];
#else
union
{
uint32_t bitwise;
f4x2_t f4x2_array[4];
} value{0};
uint8_t l = utils::sat_convert_to_type<f4_t>(x[1] / scale);
uint8_t h = utils::sat_convert_to_type<f4_t>(x[0] / scale);
value.bitwise = (h << 4) | l;
return value.f4x2_array[0];
#endif
}
// convert vector of 32 fp32 to vector of 32 fp4 with rne
inline __host__ __device__ f4x32_t f4_convert_rne(float32_t x, float scale = 1.0f)
{
#if defined(__gfx950__)
union
{
__uint128_t bitwise;
f4x2_t f4x2_array[16];
f4x32_t f4x32_array;
} f4_values{}, tmp_values{};
// TODO: pack in a loop
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[0], x[1], scale, 0);
f4_values.f4x2_array[0] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[2], x[3], scale, 0);
f4_values.f4x2_array[1] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[4], x[5], scale, 0);
f4_values.f4x2_array[2] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[6], x[7], scale, 0);
f4_values.f4x2_array[3] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[8], x[9], scale, 0);
f4_values.f4x2_array[4] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[10], x[11], scale, 0);
f4_values.f4x2_array[5] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[12], x[13], scale, 0);
f4_values.f4x2_array[6] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[14], x[15], scale, 0);
f4_values.f4x2_array[7] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[16], x[17], scale, 0);
f4_values.f4x2_array[8] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[18], x[19], scale, 0);
f4_values.f4x2_array[9] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[20], x[21], scale, 0);
f4_values.f4x2_array[10] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[22], x[23], scale, 0);
f4_values.f4x2_array[11] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[24], x[25], scale, 0);
f4_values.f4x2_array[12] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[26], x[27], scale, 0);
f4_values.f4x2_array[13] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[28], x[29], scale, 0);
f4_values.f4x2_array[14] = tmp_values.f4x2_array[0];
tmp_values.bitwise =
__builtin_amdgcn_cvt_scalef32_pk_fp4_f32(tmp_values.bitwise, x[30], x[31], scale, 0);
f4_values.f4x2_array[15] = tmp_values.f4x2_array[0];
return f4_values.f4x32_array;
#else
union
{
__uint128_t bitwise;
f4x2_t f4x2_array[16];
f4x32_t f4x32_array;
} f4_values{};
// TODO: pack in a loop
auto tmp = utils::sat_convert_to_type<f4_t>(x[0] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[1] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[2] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[3] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[4] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[5] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[6] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[7] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[8] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[9] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[10] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[11] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[12] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[13] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[14] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[15] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[16] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[17] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[18] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[19] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[20] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[21] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[22] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[23] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[24] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[25] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[26] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[27] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[28] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[29] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[30] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type<f4_t>(x[31] / scale);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
return f4_values.f4x32_array;
#endif
}
// convert fp32 to fp4 with stochastic rounding
inline __host__ __device__ f4_t f4_convert_sr(float x, float scale = 1.0f)
{
constexpr int seed = 1254739;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx950__)
union
{
uint32_t bitwise;
f4_t f4_array[4];
} value{0};
union
{
float float_array[2];
float2_t float2_array;
} float_values{{x}};
value.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
value.bitwise, float_values.float2_array, rng, scale, 0);
return value.f4_array[0];
#else
return utils::sat_convert_to_type_sr<f4_t>(x / scale, rng);
#endif
}
// convert vector of 2 fp32 to vector of 2 fp4 with sr
inline __host__ __device__ f4x2_t f4_convert_sr(float2_t x, float scale = 1.0f)
{
constexpr int seed = 1254739;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x[0]);
#if defined(__gfx950__)
union
{
uint32_t bitwise;
f4x2_t f4x2_array[4];
} value{0};
value.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(value.bitwise, x, rng, scale, 0);
return value.f4x2_array[0];
#else
union
{
uint32_t bitwise;
f4x2_t f4x2_array[4];
} value{0};
uint8_t l = utils::sat_convert_to_type_sr<f4_t>(x[1] / scale, rng);
uint8_t h = utils::sat_convert_to_type_sr<f4_t>(x[0] / scale, rng);
value.bitwise = (h << 4) | l;
return value.f4x2_array[0];
#endif
}
// convert vector of 32 fp32 to vector of 32 fp4 with sr
inline __host__ __device__ f4x32_t f4_convert_sr(float32_t x, float scale = 1.0f)
{
constexpr int seed = 1254739;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x[0]);
#if defined(__gfx950__)
union
{
__uint128_t bitwise;
f4x2_t f4x2_array[16];
f4x32_t f4x32_array;
} f4_values{0}, tmp_values{0};
union
{
float2_t floatx2_array[16];
float32_t floatx32_array;
} float_values{{0}};
// TODO: pack in a loop
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[0], rng, scale, 0);
f4_values.f4x2_array[0] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[1], rng, scale, 0);
f4_values.f4x2_array[1] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[2], rng, scale, 0);
f4_values.f4x2_array[2] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[3], rng, scale, 0);
f4_values.f4x2_array[3] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[4], rng, scale, 0);
f4_values.f4x2_array[4] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[5], rng, scale, 0);
f4_values.f4x2_array[5] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[6], rng, scale, 0);
f4_values.f4x2_array[6] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[7], rng, scale, 0);
f4_values.f4x2_array[7] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[8], rng, scale, 0);
f4_values.f4x2_array[8] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[9], rng, scale, 0);
f4_values.f4x2_array[9] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[10], rng, scale, 0);
f4_values.f4x2_array[10] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[11], rng, scale, 0);
f4_values.f4x2_array[11] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[12], rng, scale, 0);
f4_values.f4x2_array[12] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[13], rng, scale, 0);
f4_values.f4x2_array[13] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[14], rng, scale, 0);
f4_values.f4x2_array[14] = tmp_values.f4x2_array[0];
tmp_values.bitwise = __builtin_amdgcn_cvt_scalef32_sr_pk_fp4_f32(
tmp_values.bitwise, float_values.floatx2_array[15], rng, scale, 0);
f4_values.f4x2_array[15] = tmp_values.f4x2_array[0];
return f4_values.f4x32_array;
#else
union
{
__uint128_t bitwise;
f4x2_t f4x2_array[16];
f4x32_t f4x32_array;
} f4_values{0};
// TODO: pack in a loop
auto tmp = utils::sat_convert_to_type_sr<f4_t>(x[0] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[1] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[2] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[3] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[4] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[5] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[6] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[7] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[8] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[9] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[10] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[11] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[12] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[13] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[14] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[15] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[16] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[17] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[18] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[19] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[20] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[21] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[22] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[23] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[24] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[25] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[26] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[27] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[28] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[29] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[30] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
tmp = utils::sat_convert_to_type_sr<f4_t>(x[31] / scale, rng);
f4_values.bitwise <<= 4;
f4_values.bitwise |= tmp;
return f4_values.f4x32_array;
#endif
}
// convert fp32 to fp4
template <>
inline __host__ __device__ f4_t type_convert<f4_t, float>(float x)
{
#if CK_USE_SR_F4_CONVERSION
return f4_convert_sr(x);
#else
return f4_convert_rne(x);
#endif
}
// convert vector of 2 fp32 to vector of 2 fp4
template <>
inline __host__ __device__ f4x2_t type_convert<f4x2_t, float2_t>(float2_t x)
{
#if CK_USE_SR_F4_CONVERSION
return f4_convert_sr(x);
#else
return f4_convert_rne(x);
#endif
}
// convert vector of 32 fp32 to vector of 32 fp4
template <>
inline __host__ __device__ f4x32_t type_convert<f4x32_t, float32_t>(float32_t x)
{
#if CK_USE_SR_F4_CONVERSION
return f4_convert_sr(x);
#else
return f4_convert_rne(x);
#endif
}
// convert fp4 to fp32
template <>
inline __host__ __device__ float type_convert<float, f4_t>(f4_t x)
{
#if defined(__gfx950__)
union
{
float float_array[2];
float2_t float2_array;
} float_values{};
float scale = 1.0f;
float_values.float2_array = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(x, scale, 0);
return float_values.float_array[0];
#else
return utils::to_float<f4_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), x);
#endif
}
// convert vector of 2 fp4 to vector of 2 fp32
template <>
inline __host__ __device__ float2_t type_convert<float2_t, f4x2_t>(f4x2_t x)
{
#if defined(__gfx950__)
union
{
uint32_t bitwise;
f4x2_t f4x2_array[4];
} value{};
value.f4x2_array[0] = x;
float scale = 1.0f;
return __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(value.bitwise, scale, 0);
#else
float2_t ret{
utils::to_float<f4_t>(NumericLimits<e8m0_bexp_t>::Binary_1(),
x.template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{})),
utils::to_float<f4_t>(NumericLimits<e8m0_bexp_t>::Binary_1(),
x.template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}))};
return ret;
#endif
}
// convert vector of 32 fp4 to vector of 32 fp32
template <>
inline __host__ __device__ float32_t type_convert<float32_t, f4x32_t>(f4x32_t x)
{
#if defined(__gfx950__)
union
{
f4x32_t f4x32_array;
f4x2_t fp4x2[16];
} value{x};
union
{
uint32_t bitwise;
f4x2_t f4x2_array[4];
} bitwise_value{};
float2_t op;
float32_t ret;
float scale = 1.0f;
// TODO: pack in a loop
bitwise_value.f4x2_array[0] = value.fp4x2[0];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[0] = op[0];
ret[1] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[1];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[2] = op[0];
ret[3] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[2];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[4] = op[0];
ret[5] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[3];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[6] = op[0];
ret[7] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[4];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[8] = op[0];
ret[9] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[5];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[10] = op[0];
ret[11] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[6];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[12] = op[0];
ret[13] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[7];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[14] = op[0];
ret[15] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[8];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[16] = op[0];
ret[17] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[9];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[18] = op[0];
ret[19] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[10];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[20] = op[0];
ret[21] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[11];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[22] = op[0];
ret[23] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[12];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[24] = op[0];
ret[25] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[13];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[26] = op[0];
ret[27] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[14];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[28] = op[0];
ret[29] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[15];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[30] = op[0];
ret[31] = op[1];
return ret;
#else
union
{
float32_t float32_array;
float float_array[32];
} float_values{};
union
{
__uint128_t bitwise;
f4x2_t f4x2_array[16];
f4x32_t f4x32_array;
} f4_values{bit_cast<__uint128_t>(x)};
// TODO: pack in a loop
float_values.float_array[0] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[0].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[0].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[1].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[1].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[2].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[2].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[3].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[3].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[0] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[4].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[4].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[5].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[5].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[6].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[6].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[7].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[7].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[0] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[8].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[8].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[9].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[9].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[10].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[10].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[11].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[11].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[0] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[12].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[12].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[13].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[13].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[14].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[14].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[15].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
NumericLimits<e8m0_bexp_t>::Binary_1(),
f4_values.f4x2_array[15].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
return float_values.float32_array;
#endif
}
/**
* @brief Converts a float to a 6-bit float type (f6_t) using round-to-nearest-even.
*
* Divides the input by the specified scale, then saturates and converts it
* to the 6-bit floating-point format (f6_t).
*
* @param x The input float value.
* @param scale A scaling factor applied to `x` before conversion.
* @return The converted f6_t value.
*/
inline __host__ __device__ f6_t f6_convert_rne(float x, float scale = 1.0f)
{
#if defined(__gfx950__)
float16_t in1{x};
float16_t in2{};
union
{
f6x32_t f6_vector;
f6_t f6_array[32];
} out{};
out.f6_vector = __builtin_amdgcn_cvt_scalef32_2xpk16_fp6_f32(in1, in2, scale);
return out.f6_array[0];
#else
return utils::sat_convert_to_type<f6_t>(x / scale);
#endif
}
/**
* @brief Converts a 32-element single-precision float array into a packed 6-bit representation.
*
* This function divides each input float by the provided scale value, then performs conversion with
* rounding to nearest / even to pack each element into 6 bits of precision.
*
* @param x A vector of 32 floats stored in float32_t.
* @param scale A scaling factor for each float before conversion.
* @return An f6x32_t object storing the compressed 6-bit representation.
*/
inline __host__ __device__ f6x32_t f6_convert_rne(float32_t x, float scale = 1.0f)
{
#if defined(__gfx950__)
float16_t* in1 = reinterpret_cast<float16_t*>(&x);
float16_t* in2 = reinterpret_cast<float16_t*>(&x + 16);
return __builtin_amdgcn_cvt_scalef32_2xpk16_fp6_f32(*in1, *in2, scale);
#else
union
{
float32_t float_vector;
float float_array[32];
} in{x};
union
{
f6x32_t f6_vector;
f6_t f6_array[32];
} out{};
ck::static_for<0, 32, 1>{}([&](auto i) {
out.f6_array[i] = utils::sat_convert_to_type<f6_t>(in.float_array[i] / scale);
});
return out.f6_vector;
#endif
}
/**
* @brief Converts a float to the 6-bit floating-point type (f6_t) using stochastic rounding.
*
* Divides the input by the specified scale, then performs saturation and conversion
* to f6_t based on a pseudo-randomly generated seed.
*
* @param x The input float value.
* @param scale A scaling factor applied to `x` before conversion.
* @return The converted f6_t value.
*/
inline __host__ __device__ f6_t f6_convert_sr(float x, float scale = 1.0f)
{
constexpr int seed = 1254739;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx950__)
union
{
float32_t float_vector;
float float_array[32];
} in{x};
union
{
f6x32_t f6_vector;
f6_t f6_array[32];
} out{};
out.f6_vector = __builtin_amdgcn_cvt_scalef32_sr_pk32_fp6_f32(in.float_vector, rng, scale);
return out.f6_array[0];
#else
return utils::sat_convert_to_type_sr<f6_t>(x / scale, rng);
#endif
}
/**
* @brief Converts a 32-element single-precision float array into a packed 6-bit representation.
*
* This function divides each input float by the provided scale value, then performs conversion with
* stochastic rounding to pack each element into 6 bits of precision.
*
* @param x A vector of 32 floats stored in float32_t.
* @param scale A scaling factor for each float before conversion.
* @return An f6x32_t object storing the compressed 6-bit representation.
*/
inline __host__ __device__ f6x32_t f6_convert_sr(float32_t x, float scale = 1.0f)
{
constexpr int seed = 1254739;
union
{
float32_t float_vector;
float float_array[32];
} float_values{x};
uint32_t rng =
prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), float_values.float_array[0]);
#if defined(__gfx950__)
return __builtin_amdgcn_cvt_scalef32_sr_pk32_fp6_f32(x, rng, scale);
#else
union
{
float32_t float_vector;
float float_array[32];
} in{x};
union
{
f6x32_t f6_vector;
f6_t f6_array[32];
} out{};
ck::static_for<0, 32, 1>{}([&](auto i) {
out.f6_array[i] = utils::sat_convert_to_type_sr<f6_t>(in.float_array[i] / scale, rng);
});
return out.f6_vector;
#endif
}
/**
* @brief Specializes the type conversion template for converting a float into the 6-bit float type
* (f6_t).
*
* Depending on the CK_USE_SR_F6_CONVERSION flag,
* the conversion uses stochastic rounding
* or round-to-nearest-even.
*
* @param x Input float value to be converted.
* @return The converted f6_t value.
*/
template <>
inline __host__ __device__ f6_t type_convert<f6_t, float>(float x)
{
#if CK_USE_SR_F6_CONVERSION
return f6_convert_sr(x);
#else
return f6_convert_rne(x);
#endif
}
/**
* @brief Specializes the type conversion template for converting a vector of 32 floats into the
* vector of 32 6-bit float types (f6x32_t).
*
* Depending on the CK_USE_SR_F6_CONVERSION flag,
* the conversion uses stochastic rounding
* or round-to-nearest-even.
*
* @param x Input float value to be converted.
* @return The converted f6x32_t vector.
*/
template <>
inline __host__ __device__ f6x32_t type_convert<f6x32_t, float32_t>(float32_t x)
{
#if CK_USE_SR_F6_CONVERSION
return f6_convert_sr(x);
#else
return f6_convert_rne(x);
#endif
}
/**
* @brief Specializes the type conversion template for converting the 6-bit float type (f6_t) to
* float.
*
* Interprets an f6_t value as a float using the default scale factor of 1.
*
* @param x The 6-bit float (f6_t) value to be converted.
* @return The corresponding float representation.
*/
template <>
inline __host__ __device__ float type_convert<float, f6_t>(f6_t x)
{
#if defined(__gfx950__)
union
{
f6x32_t f6_vector;
f6_t f6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
out.float_vector = __builtin_amdgcn_cvt_scalef32_pk32_f32_fp6(
in.f6_vector, type_convert<float>(NumericLimits<e8m0_bexp_t>::Binary_1()));
return out.float_array[0];
#else
return utils::to_float<f6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), x);
#endif
}
/**
* @brief Specializes the type conversion template for converting the vector of 32 6-bit float types
* (f6x32_t) to vector of 32 floats.
*
* Interprets an f6_t values as floats using the default scale factor of 1.
*
* @param x The vector of 32 6-bit float (f6x32_t) values to be converted.
* @return The corresponding float representation.
*/
template <>
inline __host__ __device__ float32_t type_convert<float32_t, f6x32_t>(f6x32_t x)
{
#if defined(__gfx950__)
return __builtin_amdgcn_cvt_scalef32_pk32_f32_fp6(
x, type_convert<float>(NumericLimits<e8m0_bexp_t>::Binary_1()));
#else
union
{
f6x32_t f6_vector;
f6_t f6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
ck::static_for<0, 32, 1>{}([&](auto i) {
out.float_array[i] =
utils::to_float<f6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), in.f6_array[i]);
});
return out.float_vector;
#endif
}
/**
* @brief Converts a float to the 6-bit BF6 type using round-to-nearest-even.
*
* Divides the input by the specified scale, then saturates and converts
* it to a 6-bit BF6 floating-point format.
*
* @param x The float value to be converted.
* @param scale The scaling factor applied to the input before conversion.
* @return The converted bf6_t value.
*/
inline __host__ __device__ bf6_t bf6_convert_rne(float x, float scale = 1.0f)
{
#if defined(__gfx950__)
float16_t in1{x};
float16_t in2{};
union
{
bf6x32_t bf6_vector;
bf6_t bf6_array[32];
} out{};
out.bf6_vector = __builtin_amdgcn_cvt_scalef32_2xpk16_bf6_f32(in1, in2, scale);
return out.bf6_array[0];
#else
return utils::sat_convert_to_type<bf6_t>(x / scale);
#endif
}
/**
* @brief Converts a vector of 32 floats to the vector of 32 6-bit BF6 types using
* round-to-nearest-even.
*
* Divides the input by the specified scale, then saturates and converts
* it to a 6-bit BF6 floating-point format.
*
* @param x The float vector to be converted.
* @param scale The scaling factor applied to the input before conversion.
* @return The converted bf6x32_t vector.
*/
inline __host__ __device__ bf6x32_t bf6_convert_rne(float32_t x, float scale = 1.0f)
{
#if defined(__gfx950__)
float16_t* in1 = reinterpret_cast<float16_t*>(&x);
float16_t* in2 = reinterpret_cast<float16_t*>(&x + 16);
return __builtin_amdgcn_cvt_scalef32_2xpk16_bf6_f32(*in1, *in2, scale);
#else
union
{
float32_t float_vector;
float float_array[32];
} in{x};
union
{
bf6x32_t bf6_vector;
bf6_t bf6_array[32];
} out{};
ck::static_for<0, 32, 1>{}([&](auto i) {
out.bf6_array[i] = utils::sat_convert_to_type<bf6_t>(in.float_array[i] / scale);
});
return out.bf6_vector;
#endif
}
/**
* @brief Converts a float to the 6-bit BF6 type using stochastic rounding.
*
* Divides the input by the specified scale,
* and converts the result to a 6-bit BF6 floating-point
* format with stochastic rounding.
*
* @param x The float value to be converted.
* @param scale The scaling factor applied to the input before conversion.
* @return The converted bf6_t value.
*/
inline __host__ __device__ bf6_t bf6_convert_sr(float x, float scale = 1.0f)
{
constexpr int seed = 1254739;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx950__)
union
{
float32_t float_vector;
float float_array[32];
} in{x};
union
{
bf6x32_t bf6_vector;
bf6_t bf6_array[32];
} out{};
out.bf6_vector = __builtin_amdgcn_cvt_scalef32_sr_pk32_bf6_f32(in.float_vector, rng, scale);
return out.bf6_array[0];
#else
return utils::sat_convert_to_type_sr<bf6_t>(x / scale, rng);
#endif
}
/**
* @brief Converts a vector of 32 floats to the vector of 32 6-bit BF6 types using stochastic
* rounding.
*
* Divides the input by the specified scale,
* and converts the result to a 6-bit BF6 floating-point
* format with stochastic rounding.
*
* @param x The float vector to be converted.
* @param scale The scaling factor applied to the input before conversion.
* @return The converted bf6x32_t vector.
*/
inline __host__ __device__ bf6x32_t bf6_convert_sr(float32_t x, float scale = 1.0f)
{
constexpr int seed = 1254739;
union
{
float32_t float_vector;
float float_array[32];
} float_values{x};
uint32_t rng =
prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), float_values.float_array[0]);
#if defined(__gfx950__)
return __builtin_amdgcn_cvt_scalef32_sr_pk32_bf6_f32(x, rng, scale);
#else
union
{
float32_t float_vector;
float float_array[32];
} in{x};
union
{
bf6x32_t bf6_vector;
bf6_t bf6_array[32];
} out{};
ck::static_for<0, 32, 1>{}([&](auto i) {
out.bf6_array[i] = utils::sat_convert_to_type_sr<bf6_t>(in.float_array[i] / scale, rng);
});
return out.bf6_vector;
#endif
}
/**
* @brief Specializes float-to-bf6_t conversion.
*
* Uses stochastic rounding if CK_USE_SR_F6_CONVERSION is defined,
* otherwise uses round-to-nearest-even.
*
* @param x Input float value to convert.
* @return Converted bf6_t value.
*/
template <>
inline __host__ __device__ bf6_t type_convert<bf6_t, float>(float x)
{
#if CK_USE_SR_F6_CONVERSION
return bf6_convert_sr(x);
#else
return bf6_convert_rne(x);
#endif
}
/**
* @brief Specializes vector of 32 float-to-bf6_t conversion.
*
* Uses stochastic rounding if CK_USE_SR_F6_CONVERSION is defined,
* otherwise uses round-to-nearest-even.
*
* @param x Input float vector to convert.
* @return Converted bf6x32_t vector.
*/
template <>
inline __host__ __device__ bf6x32_t type_convert<bf6x32_t, float32_t>(float32_t x)
{
#if CK_USE_SR_F6_CONVERSION
return bf6_convert_sr(x);
#else
return bf6_convert_rne(x);
#endif
}
/**
* @brief Specializes the type conversion template for converting a bf6_t value to float.
*
* Interprets the bf6_t value using the default scale factor of 1 and returns
* its floating-point representation.
*
* @param x The bf6_t value to convert.
* @return The float representation of the given bf6_t value.
*/
template <>
inline __host__ __device__ float type_convert<float, bf6_t>(bf6_t x)
{
#if defined(__gfx950__)
union
{
bf6x32_t bf6_vector;
bf6_t bf6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
out.float_vector = __builtin_amdgcn_cvt_scalef32_pk32_f32_bf6(
in.bf6_vector, type_convert<float>(NumericLimits<e8m0_bexp_t>::Binary_1()));
return out.float_array[0];
#else
return utils::to_float<bf6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), x);
#endif
}
/**
* @brief Specializes the type conversion template for converting a vector of 32 bf6_t values to
* vector of 32 floats.
*
* Interprets the bf6x32_t value using the default scale factor of 1 and returns
* its floating-point representation.
*
* @param x The bf6x32_t value to convert.
* @return The float representation of the given vector.
*/
template <>
inline __host__ __device__ float32_t type_convert<float32_t, bf6x32_t>(bf6x32_t x)
{
#if defined(__gfx950__)
return __builtin_amdgcn_cvt_scalef32_pk32_f32_bf6(
x, type_convert<float>(NumericLimits<e8m0_bexp_t>::Binary_1()));
#else
union
{
bf6x32_t bf6_vector;
bf6_t bf6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
ck::static_for<0, 32, 1>{}([&](auto i) {
out.float_array[i] =
utils::to_float<bf6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), in.bf6_array[i]);
});
return out.float_vector;
#endif
}
#ifndef CK_CODE_GEN_RTC
template <typename Y, typename X, size_t NumElems>
inline __host__ __device__ void array_convert(std::array<Y, NumElems>& y,
......
include/ck_tile/core.hpp
View file @ 9ba504b6
......@@ -32,6 +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"
......
include/ck_tile/core/arch/generic_memory_space_atomic.hpp
View file @ 9ba504b6
// 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 @ 9ba504b6
......@@ -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
......
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