ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 14:59:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add conv bwd weight fp16 comp bf8 fp8 op, instances and example (#945)

Browse Source 
* Add f8 bf8 gemm example

* Add element-wise ops

* Add intrinsics

* Update reference calculation

* Add an additional type option for xdlops gemm

* Fix build process

* Add bf8 to buffer addressing

* Update blockwise op, split typeA and typeB

* Update for compatibility

* Uppdate naming to f8->fp8

* Update naming

* Format

* Update naming (#937)

* Add a client example

* Add computetypes to device and gridwise ops

* Add instances, update instance factory

* Format

* Fix a flag

* Add ckProfiler mode

* Fix typos

* Add an example

* Add bf8 generator

* add bf8 mfma; fixed type_convert for bf8

* move verfication ahead of timing

* Update reference calculation

* Fix reference

* Narrow down float init range

* Fix bf8 bf8 mfma

* Add bf8 @ fp8 mfma

* Update example

* Update instances

* Update profiler api

* Update for compatibility

* Format

* Remove extra example

* Clean up

* workaround convert

---------

Co-authored-by: Jing Zhang <jizha@amd.com>
This commit is contained in:
Rostyslav Geyyer
2023-10-04 08:19:08 -05:00
committed by GitHub
parent e921e1f08d
commit 42facfc6b7
22 changed files with 696 additions and 106 deletions
Download Patch File Download Diff File Expand all files Collapse all files

132
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View File

@@ -32,8 +32,12 @@ enum struct MfmaInstr
mfma_f64_16x16x4f64,
mfma_f32_32x32x16f8f8,
mfma_f32_16x16x32f8f8,
mfma_f32_32x32x16bf8bf8,
mfma_f32_16x16x32bf8bf8,
mfma_f32_32x32x16f8bf8,
mfma_f32_16x16x32f8bf8
mfma_f32_16x16x32f8bf8,
mfma_f32_32x32x16bf8f8,
mfma_f32_16x16x32bf8f8
};
template <MfmaInstr instr>
@@ -504,6 +508,52 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32f8f8>
};
#endif
#if defined CK_ENABLE_BF8
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16bf8bf8>
{
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_32x32x16bf8bf8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_16x16x32bf8bf8>
{
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_16x16x32bf8bf8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16f8bf8>
@@ -550,6 +600,52 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32f8bf8>
};
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16bf8f8>
{
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_32x32x16bf8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_16x16x32bf8f8>
{
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_16x16x32bf8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
#endif
template <typename base_type,
index_t MPerXdlops,
index_t NPerXdlops,
@@ -710,6 +806,20 @@ struct MfmaSelector
}
#endif
#if defined CK_ENABLE_BF8
template <>
static constexpr auto GetMfma<bf8_t, 32, 32>()
{
return MfmaInstr::mfma_f32_32x32x16bf8bf8;
}
template <>
static constexpr auto GetMfma<bf8_t, 16, 16>()
{
return MfmaInstr::mfma_f32_16x16x32bf8bf8;
}
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
static constexpr auto GetMfma<f8_t, 32, 32, bf8_t>()
@@ -724,6 +834,20 @@ struct MfmaSelector
}
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
template <>
static constexpr auto GetMfma<bf8_t, 32, 32, f8_t>()
{
return MfmaInstr::mfma_f32_32x32x16bf8f8;
}
template <>
static constexpr auto GetMfma<bf8_t, 16, 16, f8_t>()
{
return MfmaInstr::mfma_f32_16x16x32bf8f8;
}
#endif
static constexpr auto selected_mfma =
mfma_type<GetMfma<base_type, MPerXdlops, NPerXdlops, additional_type>()>{};
@@ -931,8 +1055,12 @@ struct XdlopsGemm
#if defined CK_ENABLE_FP8
|| is_same<base_type, f8_t>::value
#endif
#if defined CK_ENABLE_BF8
|| is_same<base_type, bf8_t>::value
#endif
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
|| (is_same<base_type, f8_t>::value && is_same<additional_type, bf8_t>::value)
|| (is_same<base_type, f8_t>::value && is_same<additional_type, bf8_t>::value) ||
(is_same<base_type, bf8_t>::value && is_same<additional_type, f8_t>::value)
#endif
,
"base base_type must be double, float, half, bfloat16, int8_t, f8_t or bf8_t!");