ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 21:21:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK Tile] Int8 Support on CK Tile GEMM (#2267)

Browse Source 
* updates to support int8 in 03_gemm example

* added comments, using aliases, helper functions

* test(gemm_universal): add test cases for int8 gemm pipeline

* fix(test_gemm): fix for failing test unit test for int8

* test(ck_tile): add int8 unit test for gemm universal

* refactor(gemm_universal): GPU reference verification for GEMM code improved

* style(gemm_universal): removed extra comments and did clang format

* merging recent changes to universal gemm to tile_engine

* ck tile engine integration work

* feat(tile_engine): add int8 support to tile engine ops/gemm

* feat(tile_engine): added 32 32 16 mfma instances to tile engine for int8

* style: Format code with clang-format-12

* refactor(tile_engine): address review comments

* style: removed unhelpful comments & unused variables.

* build: tile engine uses default config

* feat: add int8 support for CK_TILE GEMM

* style: added trailing commas to codegen_utils.py

* refactor: tile engine

* refactor: formatting and code review

* refactor: code formatting for python files

* fix: suppress build warning

* add support for gfx950

* refactor:KWarpTile size in gemms util

* Fix the branch and wrap up the k warp tile

* Add bf8 integration

* refactor: clang format and rebase

---------

Co-authored-by: zjli2013 <leezhengjiang@gmail.com>
Co-authored-by: AviralGoelAMD <aviral.goel@amd.com>
Co-authored-by: Khushbu Agarwal <khuagarw@amd.com>
This commit is contained in:
Thomas Ning
2025-06-25 08:20:35 -07:00
committed by GitHub
parent 6d6f4c76c1
commit e03293ebce
24 changed files with 815 additions and 301 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
include/ck_tile/ops/gemm/block/block_universal_gemm_as_bs_cr.hpp
View File

@@ -215,7 +215,7 @@ struct BlockUniversalGemmAsBsCr
using BLdsTile = decltype(make_static_distributed_tensor<ComputeDataType>(BLdsTileDistr));
ALdsTile a_warp_tile_;
ALdsTile b_warp_tile_;
BLdsTile b_warp_tile_;
// C += A * B
template <typename CBlockTensor, typename ASmemBlockWindow, typename BSmemBlockWindow>

15
include/ck_tile/ops/gemm/warp/warp_gemm.hpp
View File

@@ -282,4 +282,19 @@ using WarpGemmMfmaFp8Fp8F32M32N32K16SwizzleBTransposedCDistribution =
2,
swizzle_factor>>;
// int8
using WarpGemmMfma_i32_32x32x16_i8_i8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_i32_32x32x16_i8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_i32_32x32x16_i8_i8_CTransposed =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_i32_32x32x16_i8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_i32_16x16x32_i8_i8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_i32_16x16x32_i8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_i32_16x16x32_i8_i8_CTransposed =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_i32_16x16x32_i8<WGAttrCtlEnum::Default_>>>;
} // namespace ck_tile

181
include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma_impl.hpp
View File

@@ -1578,8 +1578,8 @@ struct WarpGemmAttributeMfmaImpl_i32_32x32x16_i8
DISPATCH_MFMA_CTRL_("v_mfma_i32_32x32x16_i8", Ctrl)
else
{
#if defined(__gfx94__)
c_vec = __builtin_amdgcn_mfma_i32_32x32x8i8(
#if defined(__gfx94__) or defined(__gfx95__)
c_vec = __builtin_amdgcn_mfma_i32_32x32x16_i8(
bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
#elif defined(__gfx908__) || defined(__gfx90a__)
static_for<0, 8, 1>{}([&](auto k) {
@@ -1609,6 +1609,183 @@ struct WarpGemmAttributeMfmaImpl_i32_32x32x16_i8
}
};
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
struct WarpGemmAttributeMfmaImpl_i32_16x16x32_i8
{
static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int32_t;
using AVecType = ext_vector_t<ADataType, 8>;
using BVecType = ext_vector_t<BDataType, 8>;
using CVecType = ext_vector_t<CDataType, 4>;
static constexpr index_t kM = 16;
static constexpr index_t kN = 16;
static constexpr index_t kK = 32;
static constexpr index_t kAMBlock = 1;
static constexpr index_t kBNBlock = 1;
static constexpr index_t kAMLane = 16;
static constexpr index_t kBNLane = 16;
static constexpr index_t kABKLane = 4;
static constexpr index_t kABKPerLane = 8;
static constexpr index_t kCMLane = 4;
static constexpr index_t kCNLane = 16;
static constexpr index_t kCM0PerLane = 1;
static constexpr index_t kCM1PerLane = 4; // write to 4x AccVGPRs
// c_vec += a_vec * b_vec
template <bool post_nop_ = false>
CK_TILE_DEVICE void operator()(CVecType& c_vec,
const AVecType& a_vec,
const BVecType& b_vec,
bool_constant<post_nop_> = {}) const
{
DISPATCH_MFMA_CTRL_("v_mfma_i32_16x16x32_i8", Ctrl)
else
{
#if defined(__gfx94__) or defined(__gfx95__)
c_vec = __builtin_amdgcn_mfma_i32_16x16x32_i8(
bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
#else
ck_tile::ignore = c_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
#endif
}
}
// c_vec = a_vec * b_vec
CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
{
CVecType c_vec{0};
operator()(c_vec, a_vec, b_vec);
return c_vec;
}
};
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
struct WarpGemmAttributeMfmaImpl_i32_16x16x64_i8
{
static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int32_t;
using AVecType = ext_vector_t<ADataType, 16>;
using BVecType = ext_vector_t<BDataType, 16>;
using CVecType = ext_vector_t<CDataType, 4>;
static constexpr index_t kM = 16;
static constexpr index_t kN = 16;
static constexpr index_t kK = 64;
static constexpr index_t kAMBlock = 1;
static constexpr index_t kBNBlock = 1;
static constexpr index_t kAMLane = 16;
static constexpr index_t kBNLane = 16;
static constexpr index_t kABKLane = 4;
static constexpr index_t kABKPerLane = 16;
static constexpr index_t kCMLane = 4;
static constexpr index_t kCNLane = 16;
static constexpr index_t kCM0PerLane = 1;
static constexpr index_t kCM1PerLane = 4; // write to 4x AccVGPRs
// c_vec += a_vec * b_vec
template <bool post_nop_ = false>
CK_TILE_DEVICE void operator()(CVecType& c_vec,
const AVecType& a_vec,
const BVecType& b_vec,
bool_constant<post_nop_> = {}) const
{
DISPATCH_MFMA_CTRL_("v_mfma_i32_16x16x64_i8", Ctrl)
else
{
#if defined(__gfx95__)
c_vec = __builtin_amdgcn_mfma_i32_16x16x64_i8(
bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
#else
ck_tile::ignore = c_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
#endif
}
}
// c_vec = a_vec * b_vec
CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
{
CVecType c_vec{0};
operator()(c_vec, a_vec, b_vec);
return c_vec;
}
};
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
struct WarpGemmAttributeMfmaImpl_i32_32x32x32_i8
{
static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int32_t;
using AVecType = ext_vector_t<ADataType, 16>;
using BVecType = ext_vector_t<BDataType, 16>;
using CVecType = ext_vector_t<CDataType, 16>;
static constexpr index_t kM = 32;
static constexpr index_t kN = 32;
static constexpr index_t kK = 32;
static constexpr index_t kAMBlock = 1;
static constexpr index_t kBNBlock = 1;
static constexpr index_t kAMLane = 32;
static constexpr index_t kBNLane = 32;
static constexpr index_t kABKLane = 2;
static constexpr index_t kABKPerLane = 16;
static constexpr index_t kCMLane = 2;
static constexpr index_t kCNLane = 32;
static constexpr index_t kCM0PerLane = 4;
static constexpr index_t kCM1PerLane = 4;
// c_vec += a_vec * b_vec
template <bool post_nop_ = false>
CK_TILE_DEVICE void operator()(CVecType& c_vec,
const AVecType& a_vec,
const BVecType& b_vec,
bool_constant<post_nop_> = {}) const
{
DISPATCH_MFMA_CTRL_("v_mfma_i32_32x32x32_i8", Ctrl)
else
{
#if defined(__gfx95__)
c_vec =
__builtin_amdgcn_mfma_i32_32x32x32_i8(a_vec, bit_cast<long>(b_vec), c_vec, 0, 0, 0);
#else
ck_tile::ignore = c_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
#endif
}
}
// c_vec = a_vec * b_vec
CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
{
CVecType c_vec{0};
operator()(c_vec, a_vec, b_vec);
return c_vec;
}
};
#undef DISPATCH_MFMA_
} // namespace ck_tile

17
include/ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp
View File

@@ -11,7 +11,7 @@ namespace ck_tile {
namespace impl {
template <typename AType,
typename BType,
typename CType,
typename AccType,
index_t MPerWave,
index_t NPerWave,
index_t KPerWave,
@@ -22,6 +22,7 @@ struct WarpGemmMfmaDispatcher;
// clang-format off
// fp16
// ADataType, BDataType, AccDataType, MPerWave, NPerWave, KPerWave, TransposeC, SwizzleA, UseStructuredSparsity
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32, 8, false> { using Type = WarpGemmMfmaF16F16F32M32N32K8; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32, 8, true> { using Type = WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32, 16, false> { using Type = WarpGemmMfmaF16F16F32M32N32K16; };
@@ -37,10 +38,12 @@ template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32, 16, false, true> { using Type = WarpGemmMfmaF16F16F32M32N32K16SwizzleA; };
// fp16 2:4 structural sparsity
// ADataType, BDataType, AccDataType, MPerWave, NPerWave, KPerWave, TransposeC, SwizzleA, UseStructuredSparsity
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32, 16, false, false, true> { using Type = WarpGemmSmfmacF16F16F32M32N32K16; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 16, 16, 32, false, false, true> { using Type = WarpGemmSmfmacF16F16F32M16N16K32; };
// bf16
// ADataType, BDataType, AccDataType, MPerWave, NPerWave, KPerWave, TransposeC, SwizzleA, UseStructuredSparsity
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32, 8, false> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32, 8, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32, 16, false> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16; };
@@ -56,6 +59,7 @@ template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32, 16, false, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA; };
// fp8
// ADataType, BDataType, AccDataType, MPerWave, NPerWave, KPerWave, TransposeC, SwizzleA, UseStructuredSparsity
template<> struct WarpGemmMfmaDispatcher<ck_tile::fp8_t, ck_tile::fp8_t, float, 32, 32, 16, false> { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::fp8_t, ck_tile::fp8_t, float, 32, 32, 32, false> { using Type = WarpGemmMfma_f32_32x32x32_fp8_fp8; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::fp8_t, ck_tile::fp8_t, float, 16, 16, 32, false> { using Type = WarpGemmMfma_f32_16x16x32_fp8_fp8; };
@@ -81,12 +85,19 @@ template<> struct WarpGemmMfmaDispatcher<ck_tile::fp8_t, ck_tile::bf8_t, float,
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf8_t, ck_tile::fp8_t, float, 32, 32, 64, false> { using Type = WarpGemmMfma_f32_32x32x64_bf8_fp8; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf8_t, ck_tile::bf8_t, float, 32, 32, 64, false> { using Type = WarpGemmMfma_f32_32x32x64_bf8_bf8; };
// int8
// ADataType, BDataType, AccDataType, MPerWave, NPerWave, KPerWave, TransposeC, SwizzleA, UseStructuredSparsity
template<> struct WarpGemmMfmaDispatcher<ck_tile::int8_t, ck_tile::int8_t, ck_tile::int32_t, 32, 32, 16, false> { using Type = WarpGemmMfma_i32_32x32x16_i8_i8; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::int8_t, ck_tile::int8_t, ck_tile::int32_t, 32, 32, 16, true> { using Type = WarpGemmMfma_i32_32x32x16_i8_i8_CTransposed; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::int8_t, ck_tile::int8_t, ck_tile::int32_t, 16, 16, 32, false> { using Type = WarpGemmMfma_i32_16x16x32_i8_i8; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::int8_t, ck_tile::int8_t, ck_tile::int32_t, 16, 16, 32, true> { using Type = WarpGemmMfma_i32_16x16x32_i8_i8_CTransposed; };
// clang-format on
} // namespace impl
template <typename AType,
typename BType,
typename CType,
typename AccType,
index_t MPerWave,
index_t NPerWave,
index_t KPerWave,
@@ -95,7 +106,7 @@ template <typename AType,
bool UseStructuredSparsity = false>
using WarpGemmMfmaDispatcher = typename impl::WarpGemmMfmaDispatcher<AType,
BType,
CType,
AccType,
MPerWave,
NPerWave,
KPerWave,