Skip to content

GitLab

Unverified Commit 2cc0fa26 authored by jefyang1's avatar jefyang1 Committed by GitHub
Browse files

Merge pull request #248 from ROCm/gfx950_mfma 

Gfx950 mfma
parents cde0f480 5008a960
Hide whitespace changes
Inline Side-by-side
Showing with 663 additions and 46 deletions
CMakeLists.txt
View file @ 2cc0fa26
......@@ -188,6 +188,9 @@ if (SUPPORTED_GPU_TARGETS MATCHES "gfx94" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9
message("Enabling FP8 gemms on native architectures")
add_definitions(-DCK_USE_GFX94)
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx95")
add_definitions(-DCK_USE_AMD_MFMA_GFX950)
endif()
if (SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
message("Enabling WMMA instances")
add_definitions(-DCK_USE_WMMA)
......
example/01_gemm/gemm_xdl_fp16.cpp
View file @ 2cc0fa26
......@@ -31,9 +31,7 @@ using DeviceGemmInstance0 = ck::tensor_operation::device::DeviceGemmXdl
// ######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>;
// // clang-format on
// clang-format off
using DeviceGemmInstance1 = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
// ######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// ######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
View file @ 2cc0fa26
......@@ -607,8 +607,15 @@ 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
#if defined(CK_USE_AMD_MFMA_GFX950)
// TODO: fix logic for gfx950 as it's temporary hack for passing compiling
constexpr index_t Gemm1KPack = math::max(
math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
#else
constexpr index_t Gemm1KPack =
MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
#endif
auto gemm1_blockwise_gemm = BlockwiseGemmXdlops_v2<
BlockSize,
......
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View file @ 2cc0fa26
......@@ -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>
{
......@@ -638,6 +780,95 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32bf8f8>
}
};
// TODO: fix mfma...f8f6f4 instructions
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x64f8f6f4>
{
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_32x32x64f8f6f4<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_16x16x128f8f6f4>
{
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_16x16x128f8f6f4<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_scale_f32_32x32x64f8f6f4>
{
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_scale_f32_32x32x64f8f6f4<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_scale_f32_16x16x128f8f6f4>
{
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_scale_f32_16x16x128f8f6f4<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <typename base_type,
index_t MPerXdlops,
index_t NPerXdlops,
......@@ -713,13 +944,21 @@ struct MfmaSelector
template <>
constexpr auto GetMfma<half_t, 32, 32>()
{
#if defined(CK_USE_AMD_MFMA_GFX950)
return MfmaInstr::mfma_f32_32x32x16f16;
#else
return MfmaInstr::mfma_f32_32x32x8f16;
#endif
}
template <>
constexpr auto GetMfma<half_t, 16, 16>()
{
#if defined(CK_USE_AMD_MFMA_GFX950)
return MfmaInstr::mfma_f32_16x16x32f16;
#else
return MfmaInstr::mfma_f32_16x16x16f16;
#endif
}
template <>
......@@ -743,7 +982,9 @@ struct MfmaSelector
template <>
constexpr auto GetMfma<bhalf_t, 32, 32>()
{
#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
#if defined(CK_USE_AMD_MFMA_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;
......@@ -753,14 +994,27 @@ struct MfmaSelector
template <>
constexpr auto GetMfma<bhalf_t, 16, 16>()
{
#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
#if defined(CK_USE_AMD_MFMA_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
}
#if defined(CK_USE_AMD_MFMA_GFX940)
#if defined(CK_USE_AMD_MFMA_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(CK_USE_AMD_MFMA_GFX940)
template <>
constexpr auto GetMfma<int8_t, 32, 32>()
{
......
include/ck/utility/amd_xdlops.hpp
View file @ 2cc0fa26
......@@ -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 +476,109 @@ struct intrin_mfma_f64_16x16x4f64<16, 16>
}
};
// TODO: fix ...f8f6f4 instructions
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x64f8f6f4;
template <>
struct intrin_mfma_f32_32x32x64f8f6f4<32, 32>
{
template <class FloatC>
__device__ static void Run(const f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x64_f8f6f4(
bit_cast<long>(reg_a),
bit_cast<long>(reg_b),
reg_c.template AsType<float16_t>()[Number<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 f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
reg_c.template AsType<float16_t>()(Number<0>{}) =
__builtin_amdgcn_mfma_scale_f32_32x32x64_f8f6f4(
bit_cast<long>(reg_a),
bit_cast<long>(reg_b),
reg_c.template AsType<float16_t>()[Number<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_16x16x128f8f6f4;
template <>
struct intrin_mfma_scale_f32_16x16x128f8f6f4<16, 16>
{
template <class FloatC>
__device__ static void Run(const f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
reg_c.template AsType<float4_t>()(Number<0>{}) =
__builtin_amdgcn_mfma_scale_f32_16x16x128_f8f6f4(
bit_cast<long>(reg_a),
bit_cast<long>(reg_b),
reg_c.template AsType<float4_t>()[Number<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_16x16x128f8f6f4;
template <>
struct intrin_mfma_f32_16x16x128f8f6f4<16, 16>
{
template <class FloatC>
__device__ static void Run(const f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx950__)
reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x128_f8f6f4(
bit_cast<long>(reg_a),
bit_cast<long>(reg_b),
reg_c.template AsType<float4_t>()[Number<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;
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp
View file @ 2cc0fa26
......@@ -41,13 +41,16 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_generic_instances =
std::tuple<
// clang-format off
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
template <ck::index_t NDimSpatial,
......@@ -58,11 +61,13 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances = std::tuple<
// clang-format off
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>,
......@@ -72,6 +77,7 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 8>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......@@ -84,13 +90,16 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_generic_instances =
std::tuple<
// clang-format off
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
template <ck::index_t NDimSpatial,
......@@ -101,11 +110,13 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instances = std::tuple<
// clang-format off
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>,
......@@ -115,6 +126,7 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instance
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 8>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......@@ -127,13 +139,16 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_generic_instances =
std::tuple<
// clang-format off
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
// NGCHW requires transpose, we use vector loads and stores params for them
......@@ -145,11 +160,13 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances = std::tuple<
// clang-format off
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>,
......@@ -173,6 +190,7 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......@@ -185,13 +203,16 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_generic_instances =
std::tuple<
// clang-format off
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
template <ck::index_t NDimSpatial,
......@@ -202,11 +223,13 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_instances = std::tuple<
// clang-format off
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>,
......@@ -230,6 +253,7 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_instance
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp
View file @ 2cc0fa26
......@@ -44,7 +44,9 @@ static constexpr auto ConvFwd1x1S1P0 = ConvolutionForwardSpecialization::Filter1
static constexpr auto ConvFwdOddC =
ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
#if !defined(CK_USE_AMD_MFMA_GFX950)
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
#endif // defined(CK_USE_AMD_MFMA_GFX950)
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
......@@ -56,11 +58,13 @@ template <index_t NDimSpatial,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_bf16_comp_instances = std::tuple<
// clang-format off
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
// Compute friendly
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 256, 32, 8, 8, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
......@@ -79,7 +83,7 @@ using device_grouped_conv_fwd_xdl_bf16_comp_instances = std::tuple<
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 64, 64, 8, 8, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 64, 128, 64, 8, 8, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 64, 64, 64, 8, 8, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......@@ -90,11 +94,13 @@ template <index_t NDimSpatial,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_f16_comp_instances = std::tuple<
// clang-format off
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 256, 32, 8, 8, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
......@@ -109,6 +115,7 @@ using device_grouped_conv_fwd_xdl_f16_comp_instances = std::tuple<
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 8, 8, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp
View file @ 2cc0fa26
......@@ -40,15 +40,18 @@ template <index_t NDimSpatial,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_merged_groups_bf16_instances = std::tuple<
// clang-format off
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| ACompute| BCompute| BlockGemm| NumGroups|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Type| Type| Pipeline| ToMerge|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | Scheduler| |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
// Instances with NumGroupsPerBatch > 1
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 16>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 32>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......@@ -59,15 +62,18 @@ template <index_t NDimSpatial,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_merged_groups_f16_instances = std::tuple<
// clang-format off
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
// Instances with NumGroupsPerBatch > 1
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 16>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 32>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
View file @ 2cc0fa26
......@@ -27,12 +27,15 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances = std::tuple<
// clang-format off
// clang-format off
//##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK| LoopScheduler| Pipeline|
//##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch| | |
//##########| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| Stage | | |
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#else
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
......@@ -65,6 +68,7 @@ using device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances = std::tuple<
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
View file @ 2cc0fa26
......@@ -27,12 +27,15 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances = std::tuple<
// clang-format off
// clang-format off
//##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK| LoopScheduler| Pipeline|
//##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch| | |
//##########| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| Stage | | |
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#else
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
......@@ -65,6 +68,7 @@ using device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances = std::tuple<
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
View file @ 2cc0fa26
......@@ -26,23 +26,30 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_generic_instances = std::tuple<
// clang-format off
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK| LoopScheduler| Pipeline|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch| | |
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| Stage | | |
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 64, 16, 16, 4, 16, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#else
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#endif
// clang-format on
>;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances = std::tuple<
// clang-format off
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK| LoopScheduler| Pipeline|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch| | |
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| Stage | | |
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#else
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
......@@ -102,6 +109,7 @@ using device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances = std::tuple<
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 16, 32, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
View file @ 2cc0fa26
......@@ -26,23 +26,30 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_generic_instances = std::tuple<
// clang-format off
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK| LoopScheduler| Pipeline|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch| | |
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| Stage | | |
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 4, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#else
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#endif
// clang-format on
>;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances = std::tuple<
// clang-format off
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK| LoopScheduler| Pipeline|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch| | |
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| Stage | | |
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>
#else
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v1>,
......@@ -90,6 +97,7 @@ using device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances = std::tuple<
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1, LoopScheduler::Default, PipelineVersion::v2>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
View file @ 2cc0fa26
......@@ -40,11 +40,12 @@ static constexpr auto ConvFwdOddC =
// arbitrary conv
using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##########################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -58,16 +59,20 @@ using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances = std::tuple<
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
// 1x1, pad 0
using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##########################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -81,16 +86,20 @@ using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std:
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
// 1x1, stride 1, pad 0
using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##########################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -104,15 +113,19 @@ using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = s
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_odd_c_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##########################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdOddC, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdOddC, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdOddC, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -131,6 +144,9 @@ using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_odd_c_f16_instances = std::
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdOddC, 256, 256, 64, 2, 4, 32, 32, 4, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdOddC, 128, 128, 64, 2, 4, 32, 32, 2, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdOddC, 128, 64, 64, 2, 4, 32, 32, 1, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvFwdOddC, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
View file @ 2cc0fa26
......@@ -39,11 +39,12 @@ static constexpr auto ConvFwdOddC =
// arbitrary conv
using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##########################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################################| | | | | Operation| Operation| Operation| DataOperation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -57,16 +58,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances = s
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdDefault, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdDefault, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdDefault, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdDefault, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
// 1x1, pad 0
using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##########################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################################| | | | | Operation| Operation| Operation| DataOperation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1P0, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1P0, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1P0, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -80,16 +85,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_p0_f16_instan
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1P0, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1P0, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1P0, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1P0, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
// 1x1, stride 1, pad 0
using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##########################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################################| | | | | Operation| Operation| Operation| DataOperation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1S1P0, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1S1P0, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1S1P0, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -103,16 +112,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_s1_p0_f16_ins
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1S1P0, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1S1P0, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1S1P0, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwd1x1S1P0, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
// Odd C
using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_odd_c_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##########################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################################| | | | | Operation| Operation| Operation| DataOperation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdOddC, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdOddC, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdOddC, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -131,6 +144,9 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_odd_c_f16_instanc
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdOddC, 256, 256, 64, 2, 4, 32, 32, 4, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdOddC, 128, 128, 64, 2, 4, 32, 32, 2, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdOddC, 128, 64, 64, 2, 4, 32, 32, 1, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, ConvFwdOddC, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
View file @ 2cc0fa26
......@@ -37,11 +37,12 @@ static constexpr auto ConvFwdOddC =
// arbitrary conv
using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##############################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##############################################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##############################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -55,16 +56,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdDefault, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdDefault, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdDefault, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdDefault, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
// 1x1, pad 0
using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##############################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##############################################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##############################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1P0, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1P0, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1P0, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -78,16 +83,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_p0_f16_in
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1P0, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1P0, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1P0, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1P0, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
// 1x1, stride 1, pad 0
using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##############################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##############################################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##############################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1S1P0, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1S1P0, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1S1P0, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -101,16 +110,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_s1_p0_f16
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1S1P0, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1S1P0, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1S1P0, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwd1x1S1P0, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
// Odd C
using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_odd_c_f16_instances = std::tuple<
// clang-format off
// clang-format off
//##############################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##############################################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##############################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdOddC, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdOddC, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdOddC, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
......@@ -129,6 +142,9 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_odd_c_f16_ins
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdOddC, 256, 256, 64, 2, 4, 32, 32, 4, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdOddC, 128, 128, 64, 2, 4, 32, 32, 2, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdOddC, 128, 64, 64, 2, 4, 32, 32, 1, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddReluAdd, ConvFwdOddC, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
#endif
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
View file @ 2cc0fa26
......@@ -29,12 +29,15 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances = std::tuple<
// clang-format off
// clang-format off
//#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//#####################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | Version|
//#####################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>
#else
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
......@@ -82,6 +85,7 @@ using device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances = std::tu
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, LoopScheduler::Default, PipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, LoopScheduler::Default, PipelineVersion::v2>
#endif
#endif // !defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
View file @ 2cc0fa26
......@@ -55,6 +55,11 @@ using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances = std::tuple<
DeviceGemm_Xdl_CShuffle< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemm_Xdl_CShuffleV2< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 2, 256, 256, 256, 32, 8, 8, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>
#if defined(CK_USE_AMD_MFMA_GFX950)
,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, BF16, BF16, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 64, 128, 32, 32, 16, 16, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4, LoopScheduler::Default, PipelineVersion::v1>
#endif
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
View file @ 2cc0fa26
......@@ -31,12 +31,15 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::tuple<
// clang-format off
// clang-format off
//#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//#####################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | |
//#####################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemm_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>
#else
DeviceGemm_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 2, 2, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
......@@ -93,6 +96,7 @@ using device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::tuple<
DeviceGemm_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v2>
#endif
#endif // !defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
View file @ 2cc0fa26
......@@ -31,12 +31,15 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::tuple<
// clang-format off
// clang-format off
//#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//#####################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | |
//#####################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemm_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>
#else
DeviceGemm_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
......@@ -93,6 +96,7 @@ using device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::tuple<
DeviceGemm_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v2>
#endif
#endif // !defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
View file @ 2cc0fa26
......@@ -35,22 +35,27 @@ static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecializati
static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_generic_instances = std::tuple<
// clang-format off
// clang-format off
//#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//#####################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | |
//#####################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | |
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, MNKPadding, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 8>, 1, LoopScheduler::Default, PipelineVersion::v1>
#else
DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, MNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 8>, 1, LoopScheduler::Default, PipelineVersion::v1>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
template <ck::tensor_operation::device::GemmSpecialization GemmSpec>
using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple<
// clang-format off
// clang-format off
//#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline|
//#####################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | |
//#####################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
// pipeline v1, 1 wave
DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>,
......@@ -109,6 +114,7 @@ using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple<
DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v2>,
DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v2>
#endif
#endif // !defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>;
......
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