Skip to content

GitLab

Commit c8c016dd authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of https://github.com/ROCm/composable_kernel into update_cka8w8

parents e8ca3daf 4e731776
Hide whitespace changes
Inline Side-by-side
Showing with 2319 additions and 266 deletions
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp
View file @ c8c016dd
......@@ -4,6 +4,7 @@
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp"
namespace ck_tile {
......@@ -52,6 +53,7 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
// TODO: this 8 is AK1! should be a policy parameter!
constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<kKPerBlock / 8>{}, number<kMPerBlock>{}, number<8>{}),
make_tuple(number<(kMPerBlock + 1) * 8>{}, number<8>{}, number<1>{}),
......@@ -264,6 +266,9 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
constexpr index_t M0 = MPerBlock / (M2 * M1);
static_assert(M0 * M1 * M2 == MPerBlock,
"Incorrect M0, M2, M1 configuration! "
"M0, M1, M2 must cover whole MPerBlock!");
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
......@@ -277,6 +282,9 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
{
constexpr index_t M0 = BlockSize / get_warp_size();
constexpr index_t M1 = MPerBlock / (M2 * M0);
static_assert(M0 * M1 * M2 == MPerBlock,
"Incorrect M0, M1, M2 configuration! "
"M0, M1, M2 must cover whole MPerBlock!");
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
......@@ -350,6 +358,9 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
static_assert(N2 != 0, "N2 is zero, which will lead to a division by zero error.");
static_assert(N1 != 0, "N1 is zero, which will lead to a division by zero error.");
constexpr index_t N0 = NPerBlock / (N2 * N1);
static_assert(N0 * N1 * N2 == NPerBlock,
"Incorrect N0, N1, N2 configuration! "
"N0, N1, N2 must cover whole NPerBlock!");
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
......@@ -364,7 +375,9 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
{
constexpr index_t N0 = BlockSize / get_warp_size();
constexpr index_t N1 = NPerBlock / (N2 * N0);
static_assert(N0 * N1 * N2 == NPerBlock,
"Incorrect N0, N1, N2 configuration! "
"N0, N1, N2 must cover whole NPerBlock!");
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
......@@ -475,9 +488,28 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockGemm()
{
using BlockGemmPolicy = BlockGemmASmemBSmemCRegV1DefaultPolicy;
return BlockGemmASmemBSmemCRegV1<Problem, BlockGemmPolicy>{};
constexpr bool TransposeC = false;
constexpr auto I0 = number<0>{};
constexpr auto I1 = number<1>{};
constexpr auto I2 = number<2>{};
using AccDataType = float;
using BlockWarps = typename Problem::BlockGemmShape::BlockWarps;
using WarpTile = typename Problem::BlockGemmShape::WarpTile;
using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
typename Problem::BDataType,
AccDataType,
WarpTile::at(I0),
WarpTile::at(I1),
WarpTile::at(I2),
TransposeC>;
using BlockGemmPolicy = BlockGemmASmemBSmemCRegV1CustomPolicy<typename Problem::ADataType,
typename Problem::BDataType,
typename Problem::CDataType,
BlockWarps,
WarpGemm>;
return BlockUniversalGemmAsBsCr<Problem, BlockGemmPolicy>{};
}
};
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_problem.hpp
View file @ c8c016dd
......@@ -33,6 +33,8 @@ struct GemmPipelineProblemBase
static constexpr bool kPadN = GemmTraits::kPadN;
static constexpr bool kPadK = GemmTraits::kPadK;
static constexpr auto Scheduler = GemmPipelineScheduler::Default;
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentA()
{
if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
......
include/ck_tile/ops/gemm/warp/warp_gemm.hpp
View file @ c8c016dd
......@@ -10,114 +10,134 @@
namespace ck_tile {
// fp16
using WarpGemmMfmaF16F16F32M32N32K8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M32N32K8>>;
using WarpGemmMfmaF16F16F32M16N16K16 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M16N16K16>>;
using WarpGemmMfmaF16F16F32M32N32K8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaF16F16F32M32N32K16 =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplF16F16F32M32N32K8, 2>>;
using WarpGemmMfmaF16F16F32M16N16K16 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaF16F16F32M16N16K32 =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplF16F16F32M16N16K16, 2>>;
using WarpGemmMfmaF16F16F32M32N32K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaF16F16F32M32N32K8SwizzleA = WarpGemmImpl<
WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplF16F16F32M32N32K8, 1>>;
using WarpGemmMfmaF16F16F32M16N16K32 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaF16F16F32M32N32K16SwizzleA = WarpGemmImpl<
WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplF16F16F32M32N32K8, 2>>;
using WarpGemmMfmaF16F16F32M32N32K8SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
1>>;
using WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplF16F16F32M32N32K8>>;
using WarpGemmMfmaF16F16F32M32N32K16SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplF16F16F32M16N16K16>>;
using WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8,
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M16N16K16,
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8,
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
// bf16
using WarpGemmMfmaBf16Bf16F32M32N32K8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8>>;
using WarpGemmMfmaBf16Bf16F32M16N16K16 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16>>;
using WarpGemmMfmaBf16Bf16F32M32N32K8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaBf16Bf16F32M16N16K16 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16 =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8, 2>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaBf16Bf16F32M16N16K32 =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16, 2>>;
using WarpGemmMfmaBf16Bf16F32M16N16K32 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA = WarpGemmImpl<
WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8, 1>>;
using WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
1>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA = WarpGemmImpl<
WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8, 2>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8>>;
using WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16>>;
using WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8,
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16,
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8,
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
// fp8
using WarpGemmMfma_f32_32x32x16_fp8_fp8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8>>;
using WarpGemmMfma_f32_32x32x16_fp8_bf8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8>>;
using WarpGemmMfma_f32_32x32x16_fp8_fp8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_fp8_bf8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_fp8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8>>;
using WarpGemmMfma_f32_32x32x16_bf8_fp8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_bf8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8>>;
using WarpGemmMfma_f32_32x32x16_bf8_bf8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8>>;
using WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8>>;
using WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8>>;
using WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8>>;
using WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>>>;
template <index_t swizzle_factor = 2>
using WarpGemmMfmaFp8Fp8F32M32N32K16SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, fp8_t>,
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, fp8_t, WGAttrCtlEnum::Default_>,
2,
swizzle_factor>>;
......
include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma.hpp
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -21,9 +21,12 @@ struct WarpGemmAtrributeMfma
using BVecType = typename Impl::BVecType;
using CVecType = typename Impl::CVecType;
static constexpr index_t kM = Impl::kM;
static constexpr index_t kN = Impl::kN;
static constexpr index_t kK = Impl::kK;
static constexpr index_t kM = Impl::kM;
static constexpr index_t kN = Impl::kN;
static constexpr index_t kK = Impl::kK;
static constexpr index_t kKPerThread = Impl::kABKPerLane;
CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return 1; }
using AWarpDstrEncoding = tile_distribution_encoding<
sequence<>,
......@@ -51,10 +54,13 @@ struct WarpGemmAtrributeMfma
sequence<0, 2>>;
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
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
{
Impl{}(c_vec, a_vec, b_vec);
Impl{}(c_vec, a_vec, b_vec, bool_constant<post_nop_>{});
}
// c_vec = a_vec * b_vec
......@@ -81,9 +87,12 @@ struct WarpGemmAtrributeMfmaIterateK
ext_vector_t<BDataType, vector_traits<typename Impl::BVecType>::vector_size * kKIter>;
using CVecType = typename Impl::CVecType;
static constexpr index_t kM = Impl::kM;
static constexpr index_t kN = Impl::kN;
static constexpr index_t kK = Impl::kK * kKIter;
static constexpr index_t kM = Impl::kM;
static constexpr index_t kN = Impl::kN;
static constexpr index_t kK = Impl::kK * kKIter;
static constexpr index_t kKPerThread = Impl::kABKPerLane * kKIter;
CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; }
using AWarpDstrEncoding = tile_distribution_encoding<
sequence<>,
......@@ -111,8 +120,11 @@ struct WarpGemmAtrributeMfmaIterateK
sequence<0, 2>>;
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
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
{
using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
......@@ -122,10 +134,33 @@ struct WarpGemmAtrributeMfmaIterateK
reinterpret_cast<const buf_a&>(a_vec)
.template get_as<typename Impl::AVecType>()[iKIter],
reinterpret_cast<const buf_b&>(b_vec)
.template get_as<typename Impl::BVecType>()[iKIter]);
.template get_as<typename Impl::BVecType>()[iKIter],
bool_constant<post_nop_>{});
});
}
template <index_t iKIter, bool post_nop_ = false>
CK_TILE_DEVICE void operator()(CVecType& c_vec,
const AVecType& a_vec,
const BVecType& b_vec,
number<iKIter>,
bool_constant<post_nop_> = {}) const
{
using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
static_assert(iKIter < kKIter);
// static_for<0, kKIter, 1>{}([&](auto iKIter) {
Impl{}(c_vec,
reinterpret_cast<const buf_a&>(a_vec)
.template get_as<typename Impl::AVecType>()[iKIter],
reinterpret_cast<const buf_b&>(b_vec)
.template get_as<typename Impl::BVecType>()[iKIter],
bool_constant<post_nop_>{});
//});
}
// c_vec = a_vec * b_vec
CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
{
......@@ -164,9 +199,12 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution
using BVecType = typename Impl::AVecType;
using CVecType = typename Impl::CVecType;
static constexpr index_t kM = Impl::kN;
static constexpr index_t kN = Impl::kM;
static constexpr index_t kK = Impl::kK;
static constexpr index_t kM = Impl::kN;
static constexpr index_t kN = Impl::kM;
static constexpr index_t kK = Impl::kK;
static constexpr index_t kKPerThread = Impl::kABKPerLane;
CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return 1; }
using AWarpDstrEncoding = tile_distribution_encoding<
sequence<>,
......@@ -194,11 +232,14 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution
sequence<0, 2>>;
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
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
{
// swap A and B
Impl{}(c_vec, b_vec, a_vec);
Impl{}(c_vec, b_vec, a_vec, bool_constant<post_nop_>{});
}
// c_vec = a_vec * b_vec
......@@ -222,9 +263,12 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution_SwizzleB
using BVecType = typename Impl::AVecType;
using CVecType = typename Impl::CVecType;
static constexpr index_t kM = Impl::kN;
static constexpr index_t kN = Impl::kM;
static constexpr index_t kK = Impl::kK;
static constexpr index_t kM = Impl::kN;
static constexpr index_t kN = Impl::kM;
static constexpr index_t kK = Impl::kK;
static constexpr index_t kKPerThread = Impl::kABKPerLane;
CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return 1; }
using AWarpDstrEncoding = tile_distribution_encoding<
sequence<>,
......@@ -255,12 +299,15 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution_SwizzleB
sequence<2, 2>,
sequence<0, 2>>;
template <bool post_nop_ = false>
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
CK_TILE_DEVICE void operator()(CVecType& c_vec,
const AVecType& a_vec,
const BVecType& b_vec,
bool_constant<post_nop_> = {}) const
{
// swap A and B
Impl{}(c_vec, b_vec, a_vec);
Impl{}(c_vec, b_vec, a_vec, bool_constant<post_nop_>{});
}
// c_vec = a_vec * b_vec
......@@ -287,9 +334,12 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution
ext_vector_t<BDataType, vector_traits<typename Impl::BVecType>::vector_size * kKIter>;
using CVecType = typename Impl::CVecType;
static constexpr index_t kM = Impl::kN;
static constexpr index_t kN = Impl::kM;
static constexpr index_t kK = Impl::kK * kKIter;
static constexpr index_t kM = Impl::kN;
static constexpr index_t kN = Impl::kM;
static constexpr index_t kK = Impl::kK * kKIter;
static constexpr index_t kKPerThread = Impl::kABKPerLane * kKIter;
CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; }
using AWarpDstrEncoding = tile_distribution_encoding<
sequence<>,
......@@ -316,9 +366,12 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution
sequence<2, 2>,
sequence<0, 2>>;
template <bool post_nop_ = false>
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
CK_TILE_DEVICE void operator()(CVecType& c_vec,
const AVecType& a_vec,
const BVecType& b_vec,
bool_constant<post_nop_> = {}) const
{
using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
......@@ -328,10 +381,34 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution
reinterpret_cast<const buf_b&>(b_vec)
.template get_as<typename Impl::BVecType>()[iKIter],
reinterpret_cast<const buf_a&>(a_vec)
.template get_as<typename Impl::AVecType>()[iKIter]);
.template get_as<typename Impl::AVecType>()[iKIter],
bool_constant<post_nop_>{});
});
}
template <index_t iKIter, bool post_nop_ = false>
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void operator()(CVecType& c_vec,
const AVecType& a_vec,
const BVecType& b_vec,
number<iKIter>,
bool_constant<post_nop_> = {}) const
{
using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
static_assert(iKIter < kKIter);
// swap A and B, value and type
// static_for<0, kKIter, 1>{}([&](auto iKIter) {
Impl{}(c_vec,
reinterpret_cast<const buf_b&>(b_vec)
.template get_as<typename Impl::BVecType>()[iKIter],
reinterpret_cast<const buf_a&>(a_vec)
.template get_as<typename Impl::AVecType>()[iKIter],
bool_constant<post_nop_>{});
//});
}
// c_vec = a_vec * b_vec
CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
{
......@@ -372,10 +449,13 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB
ext_vector_t<BDataType, vector_traits<typename Impl::BVecType>::vector_size * kKIter>;
using CVecType = typename Impl::CVecType;
static constexpr index_t kM = Impl::kN;
static constexpr index_t kN = Impl::kM;
static constexpr index_t kK = Impl::kK * kKIter;
static constexpr index_t SFactor = SFactor_; // group how many CM1 together
static constexpr index_t kM = Impl::kN;
static constexpr index_t kN = Impl::kM;
static constexpr index_t kK = Impl::kK * kKIter;
static constexpr index_t kKPerThread = Impl::kABKPerLane * kKIter;
static constexpr index_t SFactor = SFactor_; // group how many CM1 together
CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; }
using AWarpDstrEncoding = tile_distribution_encoding<
sequence<>,
......@@ -429,8 +509,11 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB
sequence<0, 2>>;
#endif
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
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
{
using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
......@@ -440,10 +523,33 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB
reinterpret_cast<const buf_b&>(b_vec)
.template get_as<typename Impl::BVecType>()[iKIter],
reinterpret_cast<const buf_a&>(a_vec)
.template get_as<typename Impl::AVecType>()[iKIter]);
.template get_as<typename Impl::AVecType>()[iKIter],
bool_constant<post_nop_>{});
});
}
template <index_t iKIter, bool post_nop_ = false>
CK_TILE_DEVICE void operator()(CVecType& c_vec,
const AVecType& a_vec,
const BVecType& b_vec,
number<iKIter>,
bool_constant<post_nop_> = {}) const
{
using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
static_assert(iKIter < kKIter);
// swap A and B, value and type
// static_for<0, kKIter, 1>{}([&](auto iKIter) {
Impl{}(c_vec,
reinterpret_cast<const buf_b&>(b_vec)
.template get_as<typename Impl::BVecType>()[iKIter],
reinterpret_cast<const buf_a&>(a_vec)
.template get_as<typename Impl::AVecType>()[iKIter],
bool_constant<post_nop_>{});
//});
}
// c_vec = a_vec * b_vec
CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
{
......@@ -483,10 +589,13 @@ struct WarpGemmAtrributeMfmaIterateK_SwizzleA
ext_vector_t<BDataType, vector_traits<typename Impl::BVecType>::vector_size * kKIter>;
using CVecType = typename Impl::CVecType;
static constexpr index_t kM = Impl::kM;
static constexpr index_t kN = Impl::kN;
static constexpr index_t kK = Impl::kK * kKIter;
static constexpr index_t SFactor = SFactor_; // group how many CM1 together
static constexpr index_t kM = Impl::kM;
static constexpr index_t kN = Impl::kN;
static constexpr index_t kK = Impl::kK * kKIter;
static constexpr index_t kKPerThread = Impl::kABKPerLane * kKIter;
static constexpr index_t SFactor = SFactor_; // group how many CM1 together
CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; }
using AWarpDstrEncoding = tile_distribution_encoding<
sequence<>,
......@@ -518,8 +627,11 @@ struct WarpGemmAtrributeMfmaIterateK_SwizzleA
sequence<0, 2>>;
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
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
{
using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
......@@ -529,10 +641,33 @@ struct WarpGemmAtrributeMfmaIterateK_SwizzleA
reinterpret_cast<const buf_a&>(a_vec)
.template get_as<typename Impl::AVecType>()[iKIter],
reinterpret_cast<const buf_b&>(b_vec)
.template get_as<typename Impl::BVecType>()[iKIter]);
.template get_as<typename Impl::BVecType>()[iKIter],
bool_constant<post_nop_>{});
});
}
template <index_t iKIter, bool post_nop_ = false>
CK_TILE_DEVICE void operator()(CVecType& c_vec,
const AVecType& a_vec,
const BVecType& b_vec,
number<iKIter>,
bool_constant<post_nop_> = {}) const
{
using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
static_assert(iKIter < kKIter);
// static_for<0, kKIter, 1>{}([&](auto iKIter) {
Impl{}(c_vec,
reinterpret_cast<const buf_a&>(a_vec)
.template get_as<typename Impl::AVecType>()[iKIter],
reinterpret_cast<const buf_b&>(b_vec)
.template get_as<typename Impl::BVecType>()[iKIter],
bool_constant<post_nop_>{});
//});
}
// c_vec = a_vec * b_vec
CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
{
......
include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma_impl.hpp
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -7,12 +7,68 @@
namespace ck_tile {
// TODO: refactor warp-gemm
// currently there is a discrepency for vav/vva if we need transpose C/D
// e.g. if we want A:agpr, B:vgpr, we have to use vva in WGAttrEnum
// because we swap the A/B pointer in _impl code (but not known this info here)
enum class WGAttrCtlEnum
{
Default_ = 0,
Raw_vvv = 1, // c-vgpr, a-vgpr, b-vgpr
Raw_vaa = 2, // c-vgpr, a-agpr, b-agpr
Raw_vav = 3, // c-vgpr, a-agpr, b-vgpr
Raw_vva = 4, // c-vgpr, a-vgpr, b-agpr
Raw_avv = 5, // c-agpr, a-vgpr, b-vgpr
// raw_a_a_a = 3, // c-agpr, a-agpr, b-agpr
};
#define DISPATCH_MFMA_(mfma_, dmod_, amod_, bmod_, cmod_) \
if constexpr(post_nop_) \
{ \
asm volatile(mfma_ " %0, %1, %2, %3 ; yyy\n" \
"s_nop 3" \
: dmod_(c_vec) \
: amod_(a_vec), bmod_(b_vec), cmod_(c_vec) \
:); \
} \
else \
{ \
asm volatile(mfma_ " %0, %1, %2, %3\n" \
: dmod_(c_vec) \
: amod_(a_vec), bmod_(b_vec), cmod_(c_vec) \
:); \
}
#define DISPATCH_MFMA_CTRL_(mfma_, ctrl_) \
if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vvv) \
{ \
DISPATCH_MFMA_(mfma_, "+v", "v", "v", "v") \
} \
else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vaa) \
{ \
DISPATCH_MFMA_(mfma_, "+v", "a", "a", "v") \
} \
else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vav) \
{ \
DISPATCH_MFMA_(mfma_, "+v", "a", "v", "v") \
} \
else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vva) \
{ \
DISPATCH_MFMA_(mfma_, "+v", "v", "a", "v") \
} \
else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_avv) \
{ \
DISPATCH_MFMA_(mfma_, "+a", "v", "v", "a") \
}
// FP16
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
struct WarpGemmAttributeMfmaImplF16F16F32M32N32K8
{
using ADataType = fp16_t;
using BDataType = fp16_t;
using CDataType = float;
static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
using ADataType = fp16_t;
using BDataType = fp16_t;
using CDataType = float;
using AVecType = ext_vector_t<fp16_t, 4>;
using BVecType = ext_vector_t<fp16_t, 4>;
......@@ -33,16 +89,23 @@ struct WarpGemmAttributeMfmaImplF16F16F32M32N32K8
static constexpr index_t kCM1PerLane = 4;
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
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_f32_32x32x8f16", Ctrl)
else
{
#if defined(__gfx9__)
c_vec = __builtin_amdgcn_mfma_f32_32x32x8f16(a_vec, b_vec, c_vec, 0, 0, 0);
c_vec = __builtin_amdgcn_mfma_f32_32x32x8f16(a_vec, b_vec, c_vec, 0, 0, 0);
#else
ignore = c_vec;
ignore = a_vec;
ignore = b_vec;
ck_tile::ignore = c_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
#endif
}
}
// c_vec = a_vec * b_vec
......@@ -52,18 +115,20 @@ struct WarpGemmAttributeMfmaImplF16F16F32M32N32K8
return bit_cast<CVecType>(
__builtin_amdgcn_mfma_f32_32x32x8f16(a_vec, b_vec, fp32x16_t{0.f}, 0, 0, 0));
#else
ignore = a_vec;
ignore = b_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
return CVecType{0.f};
#endif
}
};
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16
{
using ADataType = fp16_t;
using BDataType = fp16_t;
using CDataType = float;
static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
using ADataType = fp16_t;
using BDataType = fp16_t;
using CDataType = float;
using AVecType = ext_vector_t<fp16_t, 4>;
using BVecType = ext_vector_t<fp16_t, 4>;
......@@ -84,16 +149,23 @@ struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16
static constexpr index_t kCM1PerLane = 4;
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
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_f32_16x16x16f16", Ctrl)
else
{
#if defined(__gfx9__)
c_vec = __builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, c_vec, 0, 0, 0);
c_vec = __builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, c_vec, 0, 0, 0);
#else
ignore = c_vec;
ignore = a_vec;
ignore = b_vec;
ck_tile::ignore = c_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
#endif
}
}
// c_vec = a_vec * b_vec
......@@ -103,19 +175,21 @@ struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16
return bit_cast<CVecType>(
__builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, fp32x4_t{0.f}, 0, 0, 0));
#else
ignore = a_vec;
ignore = b_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
return CVecType{0.f};
#endif
}
};
// Bf16
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
struct WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8
{
using ADataType = bf16_t;
using BDataType = bf16_t;
using CDataType = float;
static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
using ADataType = bf16_t;
using BDataType = bf16_t;
using CDataType = float;
using AVecType = ext_vector_t<bf16_t, 4>;
using BVecType = ext_vector_t<bf16_t, 4>;
......@@ -136,28 +210,35 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8
static constexpr index_t kCM1PerLane = 4;
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
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_f32_32x32x8bf16_1k", Ctrl)
else
{
#if defined(__gfx90a__) || defined(__gfx94__)
c_vec = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0);
c_vec = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0);
#elif defined(__gfx908__)
static_for<0, 2, 1>{}([&](auto k) {
c_vec = __builtin_amdgcn_mfma_f32_32x32x4bf16(
reinterpret_cast<const thread_buffer<ADataType, 4>&>(a_vec)
.template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
reinterpret_cast<const thread_buffer<BDataType, 4>&>(b_vec)
.template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
c_vec,
0,
0,
0);
});
static_for<0, 2, 1>{}([&](auto k) {
c_vec = __builtin_amdgcn_mfma_f32_32x32x4bf16(
reinterpret_cast<const thread_buffer<ADataType, 4>&>(a_vec)
.template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
reinterpret_cast<const thread_buffer<BDataType, 4>&>(b_vec)
.template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
c_vec,
0,
0,
0);
});
#else
ignore = c_vec;
ignore = a_vec;
ignore = b_vec;
ck_tile::ignore = c_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
#endif
}
}
// c_vec = a_vec * b_vec
......@@ -181,18 +262,20 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8
});
return c_vec;
#else
ignore = a_vec;
ignore = b_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
return CVecType{0.f};
#endif
}
};
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
struct WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16
{
using ADataType = bf16_t;
using BDataType = bf16_t;
using CDataType = float;
static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
using ADataType = bf16_t;
using BDataType = bf16_t;
using CDataType = float;
using AVecType = ext_vector_t<bf16_t, 4>;
using BVecType = ext_vector_t<bf16_t, 4>;
......@@ -213,28 +296,34 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16
static constexpr index_t kCM1PerLane = 4;
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
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_f32_16x16x16bf16_1k", Ctrl)
{
#if defined(__gfx90a__) || defined(__gfx94__)
c_vec = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0);
c_vec = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0);
#elif defined(__gfx908__)
static_for<0, 2, 1>{}([&](auto k) {
c_vec = __builtin_amdgcn_mfma_f32_16x16x8bf16(
reinterpret_cast<const thread_buffer<ADataType, 4>&>(a_vec)
.template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
reinterpret_cast<const thread_buffer<BDataType, 4>&>(b_vec)
.template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
c_vec,
0,
0,
0);
});
static_for<0, 2, 1>{}([&](auto k) {
c_vec = __builtin_amdgcn_mfma_f32_16x16x8bf16(
reinterpret_cast<const thread_buffer<ADataType, 4>&>(a_vec)
.template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
reinterpret_cast<const thread_buffer<BDataType, 4>&>(b_vec)
.template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
c_vec,
0,
0,
0);
});
#else
ignore = c_vec;
ignore = a_vec;
ignore = b_vec;
ck_tile::ignore = c_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
#endif
}
}
// c_vec = a_vec * b_vec
......@@ -258,20 +347,21 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16
});
return c_vec;
#else
ignore = a_vec;
ignore = b_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
return CVecType{0.f};
#endif
}
};
// FP8
template <typename AType_, typename BType_>
template <typename AType_, typename BType_, WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
struct WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base
{
using ADataType = AType_;
using BDataType = BType_;
using CDataType = float;
static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
using ADataType = AType_;
using BDataType = BType_;
using CDataType = float;
using AVecType = ext_vector_t<ADataType, 8>;
using BVecType = ext_vector_t<BDataType, 8>;
......@@ -292,38 +382,120 @@ struct WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base
static constexpr index_t kCM1PerLane = 4;
// c_vec += a_vec * b_vec
CK_TILE_DEVICE void
operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
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
{
if constexpr(Ctrl == WGAttrCtlEnum::Raw_vvv)
{
if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "v", "v", "v")
}
else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "v", "v", "v")
}
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "v", "v", "v")
}
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "v", "v", "v")
}
}
else if constexpr(Ctrl == WGAttrCtlEnum::Raw_vaa)
{
if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "a", "a", "v")
}
else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "a", "a", "v")
}
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "a", "a", "v")
}
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "a", "a", "v")
}
}
else if constexpr(Ctrl == WGAttrCtlEnum::Raw_vav)
{
if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "a", "v", "v")
}
else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "a", "v", "v")
}
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "a", "v", "v")
}
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "a", "v", "v")
}
}
else if constexpr(Ctrl == WGAttrCtlEnum::Raw_vva)
{
if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "v", "a", "v")
}
else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "v", "a", "v")
}
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "v", "a", "v")
}
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
{
DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "v", "a", "v")
}
}
else
{
#if defined(__gfx94__)
if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(
bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8(
bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8(
bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8(
bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(
bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8(
bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8(
bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8(
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) {
float a_f32 =
type_convert<float>(reinterpret_cast<const thread_buffer<ADataType, 8>&>(a_vec)
.template get_as<ADataType>()[number<k>{}]);
float b_f32 =
type_convert<float>(reinterpret_cast<const thread_buffer<BDataType, 8>&>(b_vec)
.template get_as<BDataType>()[number<k>{}]);
c_vec = __builtin_amdgcn_mfma_f32_32x32x2f32(a_f32, b_f32, c_vec, 0, 0, 0);
});
static_for<0, 8, 1>{}([&](auto k) {
float a_f32 =
type_convert<float>(reinterpret_cast<const thread_buffer<ADataType, 8>&>(a_vec)
.template get_as<ADataType>()[number<k>{}]);
float b_f32 =
type_convert<float>(reinterpret_cast<const thread_buffer<BDataType, 8>&>(b_vec)
.template get_as<BDataType>()[number<k>{}]);
c_vec = __builtin_amdgcn_mfma_f32_32x32x2f32(a_f32, b_f32, c_vec, 0, 0, 0);
});
#else
ignore = c_vec;
ignore = a_vec;
ignore = b_vec;
ck_tile::ignore = c_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
#endif
}
}
// c_vec = a_vec * b_vec
......@@ -356,20 +528,97 @@ struct WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base
});
return c_vec;
#else
ignore = a_vec;
ignore = b_vec;
ck_tile::ignore = a_vec;
ck_tile::ignore = b_vec;
return CVecType{0.f};
#endif
}
};
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
using WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8 =
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, fp8_t>;
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, fp8_t, Ctrl_>;
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
using WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8 =
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, bf8_t>;
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, bf8_t, Ctrl_>;
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
using WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8 =
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<bf8_t, fp8_t>;
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<bf8_t, fp8_t, Ctrl_>;
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
using WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8 =
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<bf8_t, bf8_t>;
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<bf8_t, bf8_t, Ctrl_>;
// int8
template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
struct WarpGemmAttributeMfmaImpl_i32_32x32x16_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, 16>;
static constexpr index_t kM = 32;
static constexpr index_t kN = 32;
static constexpr index_t kK = 16;
static constexpr index_t kAMLane = 32;
static constexpr index_t kBNLane = 32;
static constexpr index_t kABKLane = 2;
static constexpr index_t kABKPerLane = 8;
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_32x32x16_i8", Ctrl)
else
{
#if defined(__gfx94__)
c_vec = __builtin_amdgcn_mfma_i32_32x32x8i8(
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) {
float a_f32 =
type_convert<float>(reinterpret_cast<const thread_buffer<ADataType, 8>&>(a_vec)
.template get_as<ADataType>()[number<k>{}]);
float b_f32 =
type_convert<float>(reinterpret_cast<const thread_buffer<BDataType, 8>&>(b_vec)
.template get_as<BDataType>()[number<k>{}]);
c_vec = __builtin_amdgcn_mfma_f32_32x32x2f32(a_f32, b_f32, 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
include/ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -21,40 +21,40 @@ struct WarpGemmMfmaDispatcher;
// clang-format off
// fp16
template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32, 8, false> { using Type = WarpGemmMfmaF16F16F32M32N32K8; };
template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32, 8, true> { using Type = WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32, 16, false> { using Type = WarpGemmMfmaF16F16F32M32N32K16; };
template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32, 16, true> { using Type = WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 16, 16, 16, false> { using Type = WarpGemmMfmaF16F16F32M16N16K16; };
template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 16, 16, 16, true> { using Type = WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 16, 16, 32, false> { using Type = WarpGemmMfmaF16F16F32M16N16K32; };
template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 16, 16, 32, true> { using Type = WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution; };
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; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32, 16, true> { using Type = WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 16, 16, 16, false> { using Type = WarpGemmMfmaF16F16F32M16N16K16; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 16, 16, 16, true> { using Type = WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 16, 16, 32, false> { using Type = WarpGemmMfmaF16F16F32M16N16K32; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 16, 16, 32, true> { using Type = WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32, 8, false, true> { using Type = WarpGemmMfmaF16F16F32M32N32K8SwizzleA; };
template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32, 16, false, true> { using Type = WarpGemmMfmaF16F16F32M32N32K16SwizzleA; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32, 8, false, true> { using Type = WarpGemmMfmaF16F16F32M32N32K8SwizzleA; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32, 16, false, true> { using Type = WarpGemmMfmaF16F16F32M32N32K16SwizzleA; };
// bf16
template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32, 8, false> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8; };
template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32, 8, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32, 16, false> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16; };
template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32, 16, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 16, 16, 16, false> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16; };
template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 16, 16, 16, true> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 16, 16, 32, false> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32; };
template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 16, 16, 32, true> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution; };
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; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32, 16, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 16, 16, 16, false> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 16, 16, 16, true> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 16, 16, 32, false> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 16, 16, 32, true> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution; };
template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32, 8, false, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA; };
template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32, 16, false, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32, 8, false, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32, 16, false, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA; };
// fp8
template<> struct WarpGemmMfmaDispatcher<fp8_t, fp8_t, float, 32, 32, 16, false> { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8; };
template<> struct WarpGemmMfmaDispatcher<fp8_t, fp8_t, float, 32, 32, 16, true> { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed; };
template<> struct WarpGemmMfmaDispatcher<fp8_t, bf8_t, float, 32, 32, 16, false> { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8; };
template<> struct WarpGemmMfmaDispatcher<fp8_t, bf8_t, float, 32, 32, 16, true> { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed; };
template<> struct WarpGemmMfmaDispatcher<bf8_t, fp8_t, float, 32, 32, 16, false> { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8; };
template<> struct WarpGemmMfmaDispatcher<bf8_t, fp8_t, float, 32, 32, 16, true> { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed; };
template<> struct WarpGemmMfmaDispatcher<bf8_t, bf8_t, float, 32, 32, 16, false> { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8; };
template<> struct WarpGemmMfmaDispatcher<bf8_t, bf8_t, float, 32, 32, 16, true> { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed; };
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, 16, true> { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::fp8_t, ck_tile::bf8_t, float, 32, 32, 16, false> { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::fp8_t, ck_tile::bf8_t, float, 32, 32, 16, true> { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf8_t, ck_tile::fp8_t, float, 32, 32, 16, false> { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf8_t, ck_tile::fp8_t, float, 32, 32, 16, true> { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf8_t, ck_tile::bf8_t, float, 32, 32, 16, false> { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8; };
template<> struct WarpGemmMfmaDispatcher<ck_tile::bf8_t, ck_tile::bf8_t, float, 32, 32, 16, true> { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed; };
// clang-format on
} // namespace impl
......
include/ck_tile/ops/gemm/warp/warp_gemm_impl.hpp
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -14,6 +14,11 @@ struct WarpGemmImpl
static constexpr index_t kM = WarpGemmAttribute::kM;
static constexpr index_t kN = WarpGemmAttribute::kN;
static constexpr index_t kK = WarpGemmAttribute::kK;
/// @brief The number of elements in K dimension processed by single thread in wavefront.
///
/// @note Note that WarpGemm may run MFMA instruction multiple times (on different K).
/// In such situation this value reflects this fact.
static constexpr index_t kKPerThread = WarpGemmAttribute::kKPerThread;
using ADataType = typename WarpGemmAttribute::ADataType;
using BDataType = typename WarpGemmAttribute::BDataType;
......@@ -31,11 +36,21 @@ struct WarpGemmImpl
using BWarpTensor = static_distributed_tensor<BDataType, BWarpDstr>;
using CWarpTensor = static_distributed_tensor<CDataType, CWarpDstr>;
CK_TILE_DEVICE void operator()(CWarpTensor& c, const AWarpTensor& a, const BWarpTensor& b) const
CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access()
{
using AVec = ext_vector_t<ADataType, AWarpTensor::get_thread_buffer_size()>;
using BVec = ext_vector_t<BDataType, BWarpTensor::get_thread_buffer_size()>;
using CVec = ext_vector_t<CDataType, CWarpTensor::get_thread_buffer_size()>;
return WarpGemmAttribute_::get_num_of_access();
}
template <typename CTensor, typename ATensor, typename BTensor, bool post_nop_ = false>
CK_TILE_DEVICE void
operator()(CTensor& c, const ATensor& a, const BTensor& b, bool_constant<post_nop_> = {}) const
{
static_assert(detail::is_similiar_distributed_tensor_v<CTensor, CWarpTensor> &&
detail::is_similiar_distributed_tensor_v<ATensor, AWarpTensor> &&
detail::is_similiar_distributed_tensor_v<BTensor, BWarpTensor>);
using AVec = ext_vector_t<ADataType, ATensor::get_thread_buffer_size()>;
using BVec = ext_vector_t<BDataType, BTensor::get_thread_buffer_size()>;
using CVec = ext_vector_t<CDataType, CTensor::get_thread_buffer_size()>;
constexpr auto I0 = number<0>{};
......@@ -44,18 +59,49 @@ struct WarpGemmImpl
auto c_vec = c.get_thread_buffer().template get_as<CVec>()[I0];
// c_vec += a_vec * b_vec
WarpGemmAttribute{}(c_vec, a_vec, b_vec);
WarpGemmAttribute{}(c_vec, a_vec, b_vec, bool_constant<post_nop_>{});
c.get_thread_buffer().template set_as<CVec>(I0, c_vec);
}
CK_TILE_DEVICE auto operator()(const AWarpTensor& a, const BWarpTensor& b) const
template <typename CTensor,
typename ATensor,
typename BTensor,
index_t i_subk,
bool post_nop_ = false>
CK_TILE_DEVICE void operator()(CTensor& c,
const ATensor& a,
const BTensor& b,
number<i_subk>,
bool_constant<post_nop_> = {}) const
{
CWarpTensor c;
using AVec = ext_vector_t<ADataType, ATensor::get_thread_buffer_size()>;
using BVec = ext_vector_t<BDataType, BTensor::get_thread_buffer_size()>;
using CVec = ext_vector_t<CDataType, CTensor::get_thread_buffer_size()>;
constexpr auto I0 = number<0>{};
using AVec = ext_vector_t<ADataType, AWarpTensor::get_thread_buffer_size()>;
using BVec = ext_vector_t<BDataType, BWarpTensor::get_thread_buffer_size()>;
using CVec = ext_vector_t<CDataType, CWarpTensor::get_thread_buffer_size()>;
const auto a_vec = a.get_thread_buffer().template get_as<AVec>()[I0];
const auto b_vec = b.get_thread_buffer().template get_as<BVec>()[I0];
auto c_vec = c.get_thread_buffer().template get_as<CVec>()[I0];
// c_vec += a_vec * b_vec
WarpGemmAttribute{}(c_vec, a_vec, b_vec, number<i_subk>{}, bool_constant<post_nop_>{});
c.get_thread_buffer().template set_as<CVec>(I0, c_vec);
}
template <typename ATensor, typename BTensor>
CK_TILE_DEVICE auto operator()(const ATensor& a, const BTensor& b) const
{
using CTensor = CWarpTensor;
static_assert(detail::is_similiar_distributed_tensor_v<ATensor, AWarpTensor> &&
detail::is_similiar_distributed_tensor_v<BTensor, BWarpTensor>);
CTensor c;
using AVec = ext_vector_t<ADataType, ATensor::get_thread_buffer_size()>;
using BVec = ext_vector_t<BDataType, BTensor::get_thread_buffer_size()>;
using CVec = ext_vector_t<CDataType, CTensor::get_thread_buffer_size()>;
constexpr auto I0 = number<0>{};
......
include/ck_tile/ops/smoothquant.hpp
View file @ c8c016dd
......@@ -3,6 +3,7 @@
#pragma once
#include "ck_tile/ops/smoothquant/kernel/moe_smoothquant_kernel.hpp"
#include "ck_tile/ops/smoothquant/kernel/smoothquant_kernel.hpp"
#include "ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_default_policy.hpp"
#include "ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_one_pass.hpp"
......
include/ck_tile/ops/smoothquant/kernel/moe_smoothquant_kernel.hpp 0 → 100644
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/common.hpp"
namespace ck_tile {
// host side args
struct MoeSmoothquantHostArgs
{
const void* p_x; // [tokens ,hidden_size], input, fp16/bf16
const void* p_xscale; // [experts, hidden_size], input, columnwise scale, fp32
const void* p_topk_ids; // [tokens, topk]
void* p_yscale; // [topk * tokens, 1], output, rowwise quant scale
void* p_qy; // [topk * tokens, hidden_size], output
index_t tokens;
index_t hidden_size;
index_t experts;
index_t topk;
index_t x_stride; // input x row stride
index_t y_stride; // output y stride(stride for topk)
};
// TODO: Extract some type to wrapper class
template <typename Pipeline_>
struct MoeSmoothquant
{
using Pipeline = remove_cvref_t<Pipeline_>;
using Problem = typename Pipeline::Problem;
using XDataType = remove_cvref_t<typename Problem::XDataType>;
using XScaleDataType = remove_cvref_t<typename Problem::XScaleDataType>;
using ComputeDataType = remove_cvref_t<typename Problem::ComputeDataType>;
using YScaleDataType = remove_cvref_t<typename Problem::YScaleDataType>;
using QYDataType = remove_cvref_t<typename Problem::QYDataType>;
static constexpr index_t Block_M = Problem::BlockShape::Block_M;
static constexpr index_t Block_N = Problem::BlockShape::Block_N;
static constexpr bool kPadM = false; // always no need to pad along M
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kTwoPass = Problem::kTwoPass;
static constexpr index_t ThreadPerWarp_N = Problem::BlockShape::ThreadPerWarp_N;
static constexpr index_t Vector_N = Problem::BlockShape::Vector_N;
static constexpr index_t Repeat_N = Problem::BlockShape::Repeat_N;
static constexpr auto I0 = number<0>{};
static constexpr auto I1 = number<1>{};
static_assert(Problem::BlockShape::Repeat_M == 1);
struct Kargs
{
const void* p_x; // [tokens ,hidden_size], input, fp16/bf16
const void* p_xscale; // [experts, hidden_size], input, columnwise scale, fp32
const void* p_topk_ids; // [tokens, topk]
void* p_yscale; // [topk, tokens, 1], output, rowwise quant scale
void* p_qy; // [topk, tokens, hidden_size], output
index_t tokens;
index_t hidden_size;
index_t experts;
index_t topk;
index_t x_stride; // input x row stride
index_t y_stride; // output y stride(stride for topk)
};
using Hargs = MoeSmoothquantHostArgs;
CK_TILE_HOST static constexpr Kargs MakeKargs(const Hargs& hargs)
{
return Kargs{hargs.p_x,
hargs.p_xscale,
hargs.p_topk_ids,
hargs.p_yscale,
hargs.p_qy,
hargs.tokens,
hargs.hidden_size,
hargs.experts,
hargs.topk,
hargs.x_stride,
hargs.y_stride};
}
CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
{
return dim3(hargs.topk, integer_divide_ceil(hargs.tokens, Block_M), 1);
}
CK_TILE_HOST static constexpr auto BlockSize() { return Problem::BlockShape::BlockSize; }
// clang-format off
template <typename T> struct t2s;
template <> struct t2s<float> { static constexpr const char * name = "fp32"; };
template <> struct t2s<ck_tile::fp16_t> { static constexpr const char * name = "fp16"; };
template <> struct t2s<ck_tile::bf16_t> { static constexpr const char * name = "bf16"; };
template <> struct t2s<ck_tile::fp8_t> { static constexpr const char * name = "fp8"; };
template <> struct t2s<ck_tile::bf8_t> { static constexpr const char * name = "bf8"; };
// clang-format on
// in byte
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize() { return Pipeline::GetSmemSize(); }
CK_TILE_HOST static std::string GetName()
{
// clang-format off
using S_ = typename Problem::BlockShape;
auto surfix = [&] () {
std::string n;
if (kPadN) n += "_pn";
if (kTwoPass) n += "_2p";
return n; }();
#define _SS_ std::string
#define _TS_ std::to_string
return _SS_("moe_smoothquant_") + _SS_(t2s<XDataType>::name) + "_" +
_TS_(S_::Block_M) + "x" + _TS_(S_::Block_N) + "_" + _TS_(S_::WarpPerBlock_M) + "x" + _TS_(S_::WarpPerBlock_N) + "_" +
_TS_(S_::Warp_M) + "x" + _TS_(S_::Warp_N) + "_" + _TS_(S_::Vector_M) + "x" + _TS_(S_::Vector_N) + "_" +
_SS_(Pipeline::name) + surfix;
#undef _SS_
#undef _TS_
// clang-format on
}
CK_TILE_DEVICE void operator()(Kargs kargs) const
{
const index_t i_topk = blockIdx.x;
const index_t i_token = blockIdx.y * Block_M;
const index_t i_token_in_thrd =
__builtin_amdgcn_readfirstlane(threadIdx.x / Problem::BlockShape::ThreadPerBlock_N);
const index_t i_expert = reinterpret_cast<const index_t*>(
kargs.p_topk_ids)[(i_token + i_token_in_thrd) * kargs.topk + i_topk];
// [tokens ,hidden_size]
const auto x_window = [&]() {
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<const XDataType*>(kargs.p_x),
make_tuple(kargs.tokens, kargs.hidden_size),
make_tuple(kargs.x_stride, 1),
number<Vector_N>{},
number<1>{});
const auto tmp2_ = pad_tensor_view(
tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<kPadM, kPadN>{});
return make_tile_window(
tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {i_token, 0});
}();
// [experts, hidden_size],
const auto xscale_window = [&]() {
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<const XScaleDataType*>(kargs.p_xscale) + i_expert * kargs.hidden_size,
make_tuple(kargs.hidden_size),
make_tuple(1),
number<Vector_N>{},
number<1>{});
const auto tmp2_ =
pad_tensor_view(tmp_, make_tuple(number<Block_N>{}), sequence<kPadN>{});
return make_tile_window(tmp2_, make_tuple(number<Block_N>{}), {0});
}();
// [topk, tokens]
auto yscale_window = [&]() {
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<YScaleDataType*>(kargs.p_yscale) + i_topk * kargs.tokens,
make_tuple(kargs.tokens),
make_tuple(1),
number<1>{});
const auto tmp2_ =
pad_tensor_view(tmp_, make_tuple(number<Block_M>{}), sequence<kPadM>{});
return make_tile_window(tmp2_, make_tuple(number<Block_M>{}), {i_token});
}();
// [topk, tokens, hidden_size]
auto qy_window = [&]() {
auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<QYDataType*>(kargs.p_qy) + i_topk * kargs.tokens * kargs.y_stride,
make_tuple(kargs.tokens, kargs.hidden_size),
make_tuple(kargs.y_stride, 1),
number<Vector_N>{},
number<1>{});
auto tmp2_ = pad_tensor_view(
tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<kPadM, kPadN>{});
return make_tile_window(
tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {i_token, 0});
}();
__shared__ char smem[GetSmemSize()];
Pipeline{}(x_window, xscale_window, yscale_window, qy_window, kargs.hidden_size, smem);
}
};
} // namespace ck_tile
include/ck_tile/ops/smoothquant/kernel/smoothquant_kernel.hpp
View file @ c8c016dd
......@@ -19,7 +19,8 @@ struct SmoothquantHostArgs
index_t m;
index_t n;
index_t stride; // row_stride
index_t x_stride; // input row_stride
index_t y_stride; // output row_stride
};
// TODO: Extract some type to wrapper class
......@@ -58,14 +59,21 @@ struct Smoothquant
index_t m;
index_t n;
index_t stride; // row_stride
index_t x_stride; // input row_stride
index_t y_stride; // out row_stride
};
using Hargs = SmoothquantHostArgs;
CK_TILE_HOST static constexpr Kargs MakeKargs(const Hargs& hargs)
{
return Kargs{
hargs.p_x, hargs.p_xscale, hargs.p_yscale, hargs.p_qy, hargs.m, hargs.n, hargs.stride};
return Kargs{hargs.p_x,
hargs.p_xscale,
hargs.p_yscale,
hargs.p_qy,
hargs.m,
hargs.n,
hargs.x_stride,
hargs.y_stride};
}
CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
......@@ -116,7 +124,7 @@ struct Smoothquant
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<const XDataType*>(kargs.p_x),
make_tuple(kargs.m, kargs.n),
make_tuple(kargs.stride, 1),
make_tuple(kargs.x_stride, 1),
number<Vector_N>{},
number<1>{});
......@@ -157,7 +165,7 @@ struct Smoothquant
auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<QYDataType*>(kargs.p_qy),
make_tuple(kargs.m, kargs.n),
make_tuple(kargs.stride, 1),
make_tuple(kargs.y_stride, 1),
number<Vector_N>{},
number<1>{});
......
include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_one_pass.hpp
View file @ c8c016dd
......@@ -25,6 +25,7 @@ struct SmoothquantPipelineOnePass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockSmoothquantProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool UseMax3 = true; // TODO - Move to trait
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
......@@ -52,7 +53,15 @@ struct SmoothquantPipelineOnePass
xscale_window_, Policy::template MakeXScaleBlockTileDistribution<Problem>());
auto reduce_absmax_func = ReduceOp::AbsMax{};
auto reduce_max_func = ReduceOp::Max{};
auto reduce_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
: "=v"(rtn)
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
auto reduce_max_func = ReduceOp::Max{};
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
auto block_reduce2d_cross_warp_sync =
......@@ -68,8 +77,23 @@ struct SmoothquantPipelineOnePass
xscale);
// compute absmax, cross-lane->cross-warp
auto absmax = block_reduce2d(
y, reduce_absmax_func.GetIdentityValue<ComputeDataType>(), reduce_absmax_func);
auto absmax = [&]() {
constexpr auto x_size_per_row =
x.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(number<1>{});
if constexpr(UseMax3 && std::is_same_v<ComputeDataType, float> &&
x_size_per_row % 2 == 0)
{
return block_reduce2d(y,
reduce_absmax_func.GetIdentityValue<ComputeDataType>(),
reduce_absmax3_func,
sequence<1, 2>{});
}
else
{
return block_reduce2d(
y, reduce_absmax_func.GetIdentityValue<ComputeDataType>(), reduce_absmax_func);
}
}();
block_reduce2d_sync(absmax, reduce_max_func);
block_reduce2d_cross_warp_sync(absmax, smem, reduce_max_func);
......
include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_two_pass.hpp
View file @ c8c016dd
......@@ -25,6 +25,7 @@ struct SmoothquantPipelineTwoPass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockSmoothquantProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool UseMax3 = true; // TODO - Move to trait
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
......@@ -56,6 +57,13 @@ struct SmoothquantPipelineTwoPass
__builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
auto reduce_absmax_func = ReduceOp::AbsMax{};
auto reduce_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
: "=v"(rtn)
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
auto reduce_max_func = ReduceOp::Max{};
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
......@@ -77,7 +85,13 @@ struct SmoothquantPipelineTwoPass
x,
xscale);
block_reduce2d(y, absmax, reduce_absmax_func);
constexpr auto x_size_per_row =
x.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(number<1>{});
if constexpr(UseMax3 && std::is_same_v<ComputeDataType, float> &&
x_size_per_row % 2 == 0)
block_reduce2d(y, absmax, reduce_absmax3_func, sequence<1, 2>{});
else
block_reduce2d(y, absmax, reduce_absmax_func);
move_tile_window(x_window, {0, Block_N});
move_tile_window(xscale_window, {Block_N});
......
include/ck_tile/ref/README.md 0 → 100644
View file @ c8c016dd
# reference
this folder contains reference implementation of a specific op. Note by including a specific header, you are including the implementation(expecially the gpu implementation) into your source code, and compile that kernel into the fatbin, hence may increase your kernel obj code length. Usually the header starts with `reference_` is a cpu reference implementation. The header starts with `naive_` contains a gpu implementation with a small launcher.
TODO: move `host/reference` under this folder
include/ck_tile/ref/naive_attention.hpp 0 → 100644
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include <thread>
#include <string>
namespace ck_tile {
enum class naive_attention_layout_enum
{
BSHD, // [batch, seqlen, nhead, hdim]
BHSD, // [batch, nhead, seqlen, hdim]
BS3HD, // [batch, nhead, 3, seqlen, hdim], used when qkv are packed
PHSD, // [pages, nhead, page_size, hdim]
// PHSDX, // [pages, nhead, page_size/x, hdim, x], where <# used pages>*page_size = seqlen
PHDSX, // [pages, nhead, hdim/x, page_size, x], where <# used pages>*page_size = seqlen
PHDS, // [pages, nhead, hdim, page_size], where <# used pages>*page_size = seqlen
};
// will used to specialize kernel variation
enum class naive_attention_variation_enum
{
FLASH_BATCHED = 0, // standard flash attention, or xformer/sdpa, used for training
FLASH_GROUPED,
DECODE_PAGED, // decode attn, where kv token from another buffer called kvcache
};
// TODO: for simplicity, this will be used as host/device arg
struct naive_attention_fwd_args
{
void* q_ptr;
void* k_ptr;
void* v_ptr;
void* o_ptr;
void* context_len_ptr; // [batch] used when seqlen kv come from a pointer(each element is a
// number, not cumsum)
void* page_table_ptr; // [batch, max_pages_per_seq] seqlen_kv is in different block(paged attn)
void* kvscale_ptr; // [nhead, 2(kv), hdim] used for kvcache dequant
float scale_s;
int hdim;
int hdim_v; // could be cross-attn, where V and Q/K hdim are different
int batch_q;
int batch_kv;
int batch_ratio_kv; // batch_q / batch_kv
int seqlen_q; // in decode case, this should be 1
int seqlen_kv; // if context_len_ptr is not nullptr, ignore this field
int nhead_q;
int nhead_kv;
int nhead_ratio_kv; // nhead_q / nhead_kv
int page_size; // if paged, the seqlen-kv per each block
int max_pages_per_seq;
};
// this is trait for host API
struct naive_attention_fwd_traits
{
std::string q_type;
std::string k_type;
std::string v_type;
std::string o_type;
std::string q_layout;
std::string k_layout;
std::string v_layout;
std::string o_layout;
int variation; // sync with naive_attention_variation_enum
};
// this is trait for kernel template
template <naive_attention_variation_enum variation_>
struct naive_attention_fwd_kernel_traits
{
static constexpr naive_attention_variation_enum variation = variation_;
};
// for simplicity, please do not use const-reference type for the template type
template <typename QType,
typename KType,
typename VType,
typename OType,
typename AccType,
naive_attention_layout_enum QLayout,
naive_attention_layout_enum KLayout,
naive_attention_layout_enum VLayout,
naive_attention_layout_enum OLayout,
typename Traits>
struct naive_attention_fwd_kernel
{
static constexpr bool is_kvcache_i8 =
std::is_same_v<KType, int8_t> && std::is_same_v<VType, int8_t> && sizeof(QType) != 1;
// kvcache-i8 will have per head scale, we apply this scale to Q/P matrix instead of original
// K/V matrix. This can speed up conversion since Q/P usually is fp16/bf16/fp32
static constexpr bool is_kvcache_i8_forward_quant = is_kvcache_i8;
// TODO: hardcode
using KVScaleType = float;
using SoftmaxType = float;
using PType = VType; // src A of gemm2, same type as V
using p_vec_type = ext_vector_t<PType, 16 / sizeof(PType)>;
static constexpr int p_vec_elem = vector_traits<p_vec_type>::vector_size;
__host__ __device__ naive_attention_fwd_kernel() {}
template <typename T, naive_attention_layout_enum Layout>
struct addresser
{
int b, s, h, d; // batch, seqlen, nhead, hdim
T* base_ptr;
__device__ addresser(int b_, int s_, int h_, int d_, void* base_ptr_)
: b(b_), s(s_), h(h_), d(d_), base_ptr(reinterpret_cast<T*>(base_ptr_))
{
}
// TODO: all the batch/nhead offset will accumulate to the base pointer
__device__ T* get_base(int i_b, int i_h)
{
if constexpr(Layout == naive_attention_layout_enum::BSHD)
return base_ptr + i_b * s * h * d + i_h * d;
else if constexpr(Layout == naive_attention_layout_enum::BHSD)
return base_ptr + i_b * s * h * d + i_h * s * d;
}
__device__ int get_offset(int i_s, int i_d)
{
if constexpr(Layout == naive_attention_layout_enum::BSHD)
return i_s * h * d + i_d;
else if constexpr(Layout == naive_attention_layout_enum::BHSD)
return i_s * d + i_d;
}
// below set of API will directly use pointer inside this struct
__device__ void init(int i_b, int i_h) { base_ptr = get_base(i_b, i_h); }
__device__ T load(int i_s, int i_d) { return base_ptr[get_offset(i_s, i_d)]; }
__device__ void store(T value, int i_s, int i_d) { base_ptr[get_offset(i_s, i_d)] = value; }
};
template <typename T, naive_attention_layout_enum Layout>
struct page_addresser
{
int s, h, d; // page_size, nhead, hdim
static constexpr int x = 16 / sizeof(T); // pack 4 dword
T* base_ptr;
int* page_table_ptr; // TODO: page table always int
int i_h; // store current head
__device__ page_addresser(int s_, int h_, int d_, void* base_ptr_, void* pptr_)
: s(s_),
h(h_),
d(d_),
base_ptr(reinterpret_cast<T*>(base_ptr_)),
page_table_ptr(reinterpret_cast<int*>(pptr_))
{
}
__device__ int64_t get_phy_page_idx(int i_s)
{
// dynamic compute page idx is simple but slow
int page_idx = i_s / s;
int phy = page_table_ptr[page_idx];
return static_cast<int64_t>(phy);
}
__device__ int get_phy_page_offset(int i_s)
{
// dynamic compute page idx is simple but slow
return i_s % s;
}
__device__ int64_t get_offset(int i_s, int i_d)
{
int page_offset = get_phy_page_offset(i_s);
int64_t page_idx = get_phy_page_idx(i_s);
int64_t base_ = page_idx * h * s * d;
if constexpr(Layout == naive_attention_layout_enum::PHSD)
return static_cast<int64_t>(i_h * s * d + page_offset * d + i_d) + base_;
else if constexpr(Layout == naive_attention_layout_enum::PHDSX)
{
int d_r = i_d / x;
int d_x = i_d % x;
return static_cast<int64_t>(i_h * d * s + d_r * s * x + page_offset * x + d_x) +
base_;
}
else if constexpr(Layout == naive_attention_layout_enum::PHDS)
{
return static_cast<int64_t>(i_h * d * s + i_d * s + page_offset) + base_;
}
}
// below set of API will directly use pointer inside this struct
__device__ void init(int /*i_b*/, int i_h_) { i_h = i_h_; }
__device__ T load(int i_s, int i_d) { return base_ptr[get_offset(i_s, i_d)]; }
__device__ void store(T /*value*/, int /*i_s*/, int /*i_d*/) {}
};
template <typename T>
struct kvscale_addresser
{
int h, d; // nhead, hdim
T* base_ptr;
__device__ kvscale_addresser(int h_, int d_, void* p_)
: h(h_), d(d_), base_ptr(reinterpret_cast<T*>(p_))
{
}
__device__ int get_offset(int i_h, int i_d, int i_kv /*0 or 1*/)
{
// [h, 2, d]
return i_h * 2 * d + i_kv * d + i_d;
}
__device__ T load(int i_h, int i_d, int i_kv)
{
return base_ptr[get_offset(i_h, i_d, i_kv)];
}
};
__device__ __host__ static constexpr int get_block_size() { return 256; }
// for simpliciy, 1 WG always compute 1 token along q, compute all token along kv
// compute all hdim from q, compute WG_SIZE hdim from v
// 1) in prefill case, seqlen_q >= 1, seqlen_kv >= 1, batch_q=batch_kv
// 2) in decode case, seqlen_q = 1, batch_q is input num-tokens, batch_kv is 1
// 3) in paged-attn case, we still use 1 WG compute all the seqlen-kv for simplicity
// TODO: could support split-kv to validate intermediate logsum
__host__ static dim3 get_grid_size(naive_attention_fwd_args args)
{
constexpr int wg_size = get_block_size();
auto g =
dim3((args.hdim_v + wg_size - 1) / wg_size, args.seqlen_q, args.batch_q * args.nhead_q);
return g;
}
// reduce single pixel within a wave
template <typename T, typename F>
__device__ constexpr T wave_reduce(T local, F reduce_f)
{
// constexpr int wave_size = 64;
constexpr int reduce_stage = 6; // 1<<6=64
T v_local = local;
#pragma unroll
for(int i_stage = 0; i_stage < reduce_stage; i_stage++)
{
int src_lane = __lane_id() ^ (1 << i_stage);
int32_t v_remote_tmp =
__builtin_amdgcn_ds_bpermute(src_lane << 2, bit_cast<int32_t>(v_local));
T v_remote = bit_cast<T>(v_remote_tmp);
v_local = reduce_f(v_local, v_remote);
}
return v_local;
}
// Note: this function must be called after wave_reduce
// Note: better not use this under if...else... with thread divergence (syncthreads)
template <typename T, typename F>
__device__ constexpr T cross_wave_reduce(T local, F reduce_f, T* smem)
{
constexpr int waves = 4;
constexpr int wave_size = 64;
int lane_id = threadIdx.x % wave_size;
__syncthreads();
smem[threadIdx.x] = local;
__syncthreads();
// the data within single wave is the same
// but for simplicity, we still use data from each lane.
T v_local = smem[lane_id];
#pragma unroll
for(int i_stage = 1; i_stage < waves; i_stage++)
{
T v_remote = smem[i_stage * wave_size + lane_id];
v_local = reduce_f(v_local, v_remote);
}
return v_local;
}
// kernel entry point
__device__ void operator()(naive_attention_fwd_args args)
{
constexpr int wg_size = get_block_size();
__shared__ char smem[wg_size * 4 * sizeof(float)]; // should enough
int i_dv = blockIdx.x * wg_size + threadIdx.x; // index of hdim_v
int i_sq = blockIdx.y; // index of seqlen_q
int i_batch = blockIdx.z; // index of batch_q * nhead_q
int i_bq = i_batch / args.nhead_q; // index of batch_q
int i_hq = i_batch % args.nhead_q; // index of nhead_q
int i_bk = i_bq / args.batch_ratio_kv;
int i_hk = i_hq / args.nhead_ratio_kv;
void* page_table_ptr = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return reinterpret_cast<int*>(args.page_table_ptr) + i_bq * args.max_pages_per_seq;
}
else
{
return nullptr;
}
}();
auto q_addr = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::FLASH_BATCHED)
{
return addresser<QType, QLayout>{
args.batch_q, args.seqlen_q, args.nhead_q, args.hdim, args.q_ptr};
}
else if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return addresser<QType, QLayout>{
args.batch_q, args.seqlen_q, args.nhead_q, args.hdim, args.q_ptr};
}
}();
auto k_addr = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::FLASH_BATCHED)
{
return addresser<KType, KLayout>{
args.batch_kv, args.seqlen_kv, args.nhead_kv, args.hdim, args.k_ptr};
}
else if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return page_addresser<KType, KLayout>{
args.page_size, args.nhead_kv, args.hdim, args.k_ptr, page_table_ptr};
}
}();
auto v_addr = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::FLASH_BATCHED)
{
return addresser<VType, VLayout>{
args.batch_kv, args.seqlen_kv, args.nhead_kv, args.hdim_v, args.v_ptr};
}
else if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return page_addresser<VType, VLayout>{
args.page_size, args.nhead_kv, args.hdim_v, args.v_ptr, page_table_ptr};
}
}();
auto o_addr = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::FLASH_BATCHED)
{
return addresser<OType, OLayout>{
args.batch_q, args.seqlen_q, args.nhead_q, args.hdim_v, args.o_ptr};
}
else if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return addresser<OType, OLayout>{
args.batch_q, args.seqlen_q, args.nhead_q, args.hdim_v, args.o_ptr};
}
}();
q_addr.init(i_bq, i_hq);
k_addr.init(i_bk, i_hk);
v_addr.init(i_bk, i_hk);
o_addr.init(i_bq, i_hq);
auto f_max = [](auto x_, auto y_) { return max(x_, y_); };
auto f_sum = [](auto x_, auto y_) { return x_ + y_; };
auto f_absmax_f32 = [](float v_0_, float v_1_) {
float rtn;
asm volatile("v_max_f32 %0, abs(%1), abs(%2)" : "=v"(rtn) : "v"(v_0_), "v"(v_1_));
return rtn;
};
int seqlen_kv = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::FLASH_BATCHED)
{
return args.seqlen_kv;
}
else if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return reinterpret_cast<int*>(args.context_len_ptr)[i_bq];
}
}();
SoftmaxType row_max = -numeric<SoftmaxType>::infinity();
SoftmaxType l{0};
AccType o_acc = {0};
int sk_loops = (seqlen_kv + wg_size - 1) / wg_size;
float qf_scale = .0f;
kvscale_addresser<KVScaleType> kvscale_addr{args.nhead_kv, args.hdim, args.kvscale_ptr};
if constexpr(is_kvcache_i8_forward_quant)
{
// AccType is i32 now, seqlen_q = 1, hdim up to 256
float q = 0;
float k_s = 0;
if(static_cast<int>(threadIdx.x) < args.hdim)
{
q = type_convert<float>(q_addr.load(0, threadIdx.x));
k_s = type_convert<float>(kvscale_addr.load(i_hk, threadIdx.x, 0));
}
// 1) we apply the k scale to q
float q_forwarded = q * k_s;
// 2) apply smooth-quant
// find absmax
float qf_max = wave_reduce(q_forwarded, f_absmax_f32);
qf_max = cross_wave_reduce(qf_max, f_absmax_f32, reinterpret_cast<float*>(smem));
// per-token scale
qf_scale = qf_max / 127.0;
// devide by scale
q = q / qf_scale;
// fp32->i8
int8_t quantized_q = static_cast<int8_t>(q);
__syncthreads();
reinterpret_cast<int8_t*>(smem)[threadIdx.x] = quantized_q;
__syncthreads();
// after above process, we have 2 data
// 1) int8 q data stored in smem(no need to reload)
// 2) per-token scale qf_scale, to be mul after 1st gemm
}
for(int i_loop1 = 0; i_loop1 < sk_loops; i_loop1++)
{
int i_sk = i_loop1 * wg_size + threadIdx.x;
// gemm-1
SoftmaxType s_softmax = -numeric<SoftmaxType>::infinity();
if(i_sk < seqlen_kv)
{
AccType s_acc{0}; // clear for every loop
for(auto i_dq = 0; i_dq < args.hdim; i_dq++)
{
if constexpr(is_kvcache_i8_forward_quant)
{
int8_t q = reinterpret_cast<int8_t*>(smem)[i_dq];
auto k = k_addr.load(i_sk, i_dq);
s_acc += type_convert<AccType>(q) * type_convert<AccType>(k);
}
else
{
auto q = q_addr.load(i_sq, i_dq); // q will have duplicate load
auto k = k_addr.load(i_sk, i_dq);
s_acc += type_convert<AccType>(q) * type_convert<AccType>(k);
}
}
// scale
s_softmax = type_convert<SoftmaxType>(s_acc);
s_softmax *=
type_convert<SoftmaxType>(args.scale_s * ck_tile::log2e_v<SoftmaxType>);
if constexpr(is_kvcache_i8_forward_quant)
{
s_softmax *= qf_scale; // post scale the per-token factor
}
}
// s->p
float pf_scale = 0.; // used for i8 quant
{
// softmax, find max
SoftmaxType old_max = row_max;
SoftmaxType cur_max = wave_reduce(s_softmax, f_max);
cur_max = cross_wave_reduce(cur_max, f_max, reinterpret_cast<SoftmaxType*>(smem));
row_max = max(old_max, cur_max); // update row_max
// softmax, exp(i_elem - max)
SoftmaxType p_compute = __builtin_amdgcn_exp2f(s_softmax - row_max);
// compute exp_sum
SoftmaxType row_sum = wave_reduce(p_compute, f_sum);
row_sum = cross_wave_reduce(row_sum, f_sum, reinterpret_cast<SoftmaxType*>(smem));
// l, pre-scall o_acc
SoftmaxType tmp = __builtin_amdgcn_exp2f(old_max - row_max);
l = tmp * l + row_sum;
o_acc = type_convert<AccType>(type_convert<SoftmaxType>(o_acc) * tmp);
// prepare the p_compute into smem, to let every thread read same p_compute and do
// 2nd gemm
if constexpr(is_kvcache_i8_forward_quant)
{
float v_s = 0;
if(static_cast<int>(threadIdx.x) < args.hdim_v)
{
v_s = type_convert<float>(kvscale_addr.load(i_hk, threadIdx.x, 1));
}
// 1) we apply the v scale to p
float p_forwarded = p_compute * v_s;
// 2) apply smooth-quant
// find absmax
float pf_max = wave_reduce(p_forwarded, f_absmax_f32);
pf_max =
cross_wave_reduce(pf_max, f_absmax_f32, reinterpret_cast<float*>(smem));
// per-token scale
pf_scale = pf_max / 127.0;
// devide by scale
p_compute = p_compute / pf_scale;
// fp32->i8
int8_t quantized_p = static_cast<int8_t>(p_compute);
__syncthreads();
reinterpret_cast<int8_t*>(smem)[threadIdx.x] = quantized_p;
__syncthreads();
// after above process, we have 2 data
// 1) int8 p data stored in smem(no need to reload)
// 2) per-token scale pf_scale, to be mul after 2nd gemm
}
else
{
__syncthreads();
reinterpret_cast<PType*>(smem)[threadIdx.x] = type_convert<PType>(p_compute);
__syncthreads();
}
}
// gemm-2, simple loop over vector by vector
constexpr int gemm_2_loop = wg_size / p_vec_elem;
{
AccType o_acc_local = {0};
int sk_start = i_loop1 * wg_size; // we start from the first seqlen_kv element
for(int i_loop2 = 0; i_loop2 < gemm_2_loop; i_loop2++)
{
p_vec_type p_vec = reinterpret_cast<p_vec_type*>(smem)[i_loop2];
#pragma unroll
for(int i_j = 0; i_j < p_vec_elem; i_j++)
{
int sv_offset = i_loop2 * p_vec_elem + i_j;
int i_sv = sk_start + sv_offset;
VType v = 0.f;
if(i_dv < args.hdim_v && i_sv < seqlen_kv)
{
v = v_addr.load(i_sv, i_dv);
}
o_acc_local += type_convert<AccType>(p_vec[i_j]) * type_convert<AccType>(v);
}
}
if constexpr(is_kvcache_i8_forward_quant)
{
// apply pr scale to local acc
o_acc_local =
type_convert<AccType>(type_convert<float>(o_acc_local) * pf_scale);
}
o_acc += o_acc_local;
}
}
// post scale o_acc
{
SoftmaxType tmp = l == 0.f ? 0.f : 1.f / l; // in case masking
o_acc = type_convert<AccType>(type_convert<SoftmaxType>(o_acc) * tmp);
}
// store O
if(i_dv < args.hdim_v)
o_addr.store(type_convert<OType>(o_acc), i_sq, i_dv);
}
};
#define CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_INTERNAL_() \
{ \
using ktraits_ = \
naive_attention_fwd_kernel_traits<static_cast<naive_attention_variation_enum>( \
variation_)>; \
using k_ = naive_attention_fwd_kernel<q_type_, \
k_type_, \
v_type_, \
o_type_, \
acc_type_, \
q_layout_, \
k_layout_, \
v_layout_, \
o_layout_, \
ktraits_>; \
dim3 grids = k_::get_grid_size(a); \
r = ck_tile::launch_kernel(s, \
ck_tile::make_kernel(k_{}, grids, k_::get_block_size(), 0, a)); \
}
#define CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_() \
if(t.variation == 0 && t.q_layout == "bshd" && t.k_layout == "bshd" && t.v_layout == "bshd" && \
t.o_layout == "bshd") \
{ \
constexpr auto q_layout_ = naive_attention_layout_enum::BSHD; \
constexpr auto k_layout_ = naive_attention_layout_enum::BSHD; \
constexpr auto v_layout_ = naive_attention_layout_enum::BSHD; \
constexpr auto o_layout_ = naive_attention_layout_enum::BSHD; \
constexpr int variation_ = 0; \
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_INTERNAL_(); \
} \
else if(t.variation == 0 && t.q_layout == "bhsd" && t.k_layout == "bhsd" && \
t.v_layout == "bhsd" && t.o_layout == "bhsd") \
{ \
constexpr auto q_layout_ = naive_attention_layout_enum::BHSD; \
constexpr auto k_layout_ = naive_attention_layout_enum::BHSD; \
constexpr auto v_layout_ = naive_attention_layout_enum::BHSD; \
constexpr auto o_layout_ = naive_attention_layout_enum::BHSD; \
constexpr int variation_ = 0; \
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_INTERNAL_(); \
} \
else if(t.variation == 2 && t.q_layout == "bhsd" && t.k_layout == "phdsx" && \
t.v_layout == "phds" && t.o_layout == "bhsd") \
{ \
constexpr auto q_layout_ = naive_attention_layout_enum::BHSD; \
constexpr auto k_layout_ = naive_attention_layout_enum::PHDSX; \
constexpr auto v_layout_ = naive_attention_layout_enum::PHDS; \
constexpr auto o_layout_ = naive_attention_layout_enum::BHSD; \
constexpr int variation_ = 2; \
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_INTERNAL_(); \
}
//
CK_TILE_HOST float naive_attention_fwd(naive_attention_fwd_traits t,
naive_attention_fwd_args a,
ck_tile::stream_config s)
{
float r = -1;
// TODO: do not explicitly create too much instance!
if(t.q_type == "fp16" && t.k_type == "fp16" && t.v_type == "fp16" && t.o_type == "fp16")
{
using q_type_ = fp16_t;
using k_type_ = fp16_t;
using v_type_ = fp16_t;
using o_type_ = fp16_t;
using acc_type_ = float;
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_();
}
else if(t.q_type == "bf16" && t.k_type == "bf16" && t.v_type == "bf16" && t.o_type == "bf16")
{
using q_type_ = bf16_t;
using k_type_ = bf16_t;
using v_type_ = bf16_t;
using o_type_ = bf16_t;
using acc_type_ = float;
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_();
}
else if(t.q_type == "bf16" && t.k_type == "int8" && t.v_type == "int8" && t.o_type == "bf16")
{
using q_type_ = bf16_t;
using k_type_ = int8_t;
using v_type_ = int8_t;
using o_type_ = bf16_t;
using acc_type_ = int32_t; // NOTE!
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_();
}
else if(t.q_type == "fp16" && t.k_type == "int8" && t.v_type == "int8" && t.o_type == "fp16")
{
using q_type_ = fp16_t;
using k_type_ = int8_t;
using v_type_ = int8_t;
using o_type_ = fp16_t;
using acc_type_ = int32_t; // NOTE!
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_();
}
return r;
}
#undef CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_
#undef CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_INTERNAL_
} // namespace ck_tile
include/ck_tile/remod.py
View file @ c8c016dd
......@@ -7,6 +7,7 @@ import copy
NS = 'ck_tile'
OPS = 'ops'
REF = 'ref'
OPS_COMMON = 'common' # common header will be duplicated into ops/* other module
HEADER_COMMON = f"""// SPDX-License-Identifier: MIT
......@@ -29,6 +30,9 @@ class submodule_t:
def push(self, f):
if len(f.parents) != 1: # ignore ./xxx.hpp
mod = get_module(f)
# ref is supposed to include one header on demand
if mod == REF:
return
if mod == OPS:
if mod not in self.m.keys():
self.m[mod] = dict()
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -62,9 +62,9 @@ struct ReferenceGemm : public device::BaseOperator
auto f_mk_kn_mn = [&](auto m, auto n) {
const int K = arg.a_m_k_.mDesc.GetLengths()[1];
AccDataType v_acc = 0;
ComputeTypeA v_a = 0;
ComputeTypeB v_b = 0;
AccDataType v_acc{0};
ComputeTypeA v_a{0};
ComputeTypeB v_b{0};
for(int k = 0; k < K; ++k)
{
......@@ -93,7 +93,7 @@ struct ReferenceGemm : public device::BaseOperator
ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
}
CDataType v_c = 0;
CDataType v_c{0};
arg.c_element_op_(v_c, v_acc);
......
library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_streamk.hpp
View file @ c8c016dd
......@@ -237,6 +237,206 @@ void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_mnkpaddin
PassThrough,
PassThrough>>>& instances);
#endif
#if(defined(CK_ENABLE_FP8))
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemm_Streamk_V2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
template <typename ADataType,
typename BDataType,
typename CDataType,
......@@ -327,6 +527,121 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemm_S
}
#endif
#if(defined(CK_ENABLE_FP8))
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_comp_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_comp_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v1_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_comp_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_comp_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v1_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f16_f8_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_comp_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_comp_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v1_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_comp_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_comp_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v1_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_streamk_f8_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
}
#endif
return op_ptrs;
}
};
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
View file @ c8c016dd
......@@ -95,6 +95,45 @@ void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_pv1_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_pv1_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_pv2_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Col,
......@@ -189,6 +228,124 @@ void add_device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_nk_mn_in
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_kn_mn_irregular_pv1_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Row,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_kn_mn_irregular_pv1_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Row,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_kn_mn_irregular_pv2_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Row,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_nk_mn_irregular_pv1_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Col,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_nk_mn_irregular_pv1_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Col,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_nk_mn_irregular_pv2_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Col,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_km_kn_mn_irregular_pv1_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
Row,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_km_kn_mn_irregular_pv1_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
Row,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_km_kn_mn_irregular_pv2_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
Row,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#if defined(CK_ENABLE_BF16) && defined(CK_ENABLE_INT8)
......@@ -262,7 +419,11 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_pv1_inter_instances(
op_ptrs);
add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_pv1_instances(
op_ptrs);
add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_pv2_instances(
op_ptrs);
add_device_grouped_gemm_multiple_d_xdl_two_stage_f16_f16_f16_mk_kn_mn_instances(
op_ptrs);
......@@ -334,12 +495,34 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_kn_mn_instances(
op_ptrs);
add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_kn_mn_irregular_pv1_inter_instances(
op_ptrs);
add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_kn_mn_irregular_pv1_instances(
op_ptrs);
add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_kn_mn_irregular_pv2_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<ELayout, Row>)
{
add_device_grouped_gemm_multiple_d_xdl_two_stage_bf16_bf16_bf16_mk_nk_mn_instances(
op_ptrs);
add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_nk_mn_irregular_pv1_inter_instances(
op_ptrs);
add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_nk_mn_irregular_pv1_instances(
op_ptrs);
add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_mk_nk_mn_irregular_pv2_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_km_kn_mn_irregular_pv1_inter_instances(
op_ptrs);
add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_km_kn_mn_irregular_pv1_instances(
op_ptrs);
add_device_grouped_gemm_xdl_splitk_bf16_bf16_bf16_km_kn_mn_irregular_pv2_instances(
op_ptrs);
}
}
#endif
......
library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp → library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_instance.hpp
View file @ c8c016dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
......@@ -7,16 +9,17 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
#include "ck/utility/loop_scheduler.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -25,97 +28,109 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Empty_Tuple = ck::Tuple<>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
static constexpr auto PipelineV1 = ck::PipelineVersion::v1;
static constexpr auto PipelineV2 = ck::PipelineVersion::v2;
static constexpr auto DefaultScheduler = ck::LoopScheduler::Default;
static constexpr auto InterwaveScheduler = ck::LoopScheduler::Interwave;
static constexpr auto GemmMNKPadding = device::GemmSpecialization::MNKPadding;
static constexpr auto GemmDefault = device::GemmSpecialization::Default;
using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
template <typename T,
device::GemmSpecialization GemmSpec = GemmMNKPadding,
PipelineVersion Pipeline = PipelineV1,
LoopScheduler Scheduler = DefaultScheduler,
enable_if_t<sizeof(T) == 2, bool> = false>
using device_grouped_gemm_xdl_splitk_2Bt_rrr_instances = std::tuple<
// clang-format off
//################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| 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|
//################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| 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|
//################################| | | | | | | | | | | 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|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 192, 64, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 48, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 24, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 192, 32, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 64, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 32, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, PipelineVersion::v1>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 192, 64, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 48, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 24, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 192, 32, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 64, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 32, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, PipelineVersion::v1, LoopScheduler::Interwave>,
//################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| 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| Pipeline | Loop |
//################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| 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| Version | Scheduler |
//################################| | | | | | | | | | | 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| | |
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 192, 64, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 48, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 32, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 24, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 192, 32, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 32, 64, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 32, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Pipeline, Scheduler>
// clang-format on
>;
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 192, 64, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 48, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 24, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 192, 32, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 64, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 32, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 128, 64, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, PipelineVersion::v2>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, PipelineVersion::v2>
template <typename T,
device::GemmSpecialization GemmSpec = GemmMNKPadding,
PipelineVersion Pipeline = PipelineV1,
LoopScheduler Scheduler = DefaultScheduler,
enable_if_t<sizeof(T) == 2, bool> = false>
using device_grouped_gemm_xdl_splitk_2Bt_rcr_instances = std::tuple<
// clang-format off
//################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| 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| Pipeline | Loop |
//################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| 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| Version | Scheduler |
//################################| | | | | | | | | | | 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| | |
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Row, Col, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Pipeline, Scheduler>
// clang-format on
>;
void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances, device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
}
template <typename T,
device::GemmSpecialization GemmSpec = GemmMNKPadding,
PipelineVersion Pipeline = PipelineV1,
LoopScheduler Scheduler = DefaultScheduler,
enable_if_t<sizeof(T) == 2, bool> = false>
using device_grouped_gemm_xdl_splitk_2Bt_crr_instances = std::tuple<
// clang-format off
//################################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| 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| Pipeline | Loop |
//################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| 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| Version | Scheduler |
//################################| | | | | | | | | | | 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| | |
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 192, 64, 32, 8, 8, 32, 32, 3, 1, S<1, 4, 48, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, S<1, 4, 48, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 32, 192, 32, 8, 8, 32, 32, 1, 3, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, S<1, 4, 24, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 192, 32, 32, 8, 8, 32, 32, 3, 1, S<1, 2, 48, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 32, 64, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 32, 32, 8, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 64, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Pipeline, Scheduler>,
DeviceGroupedGemmXdlSplitKCShuffle< Col, Row, Empty_Tuple, Row, T, T, F32, T, Empty_Tuple, T, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Pipeline, Scheduler>
// clang-format on
>;
} // namespace instance
} // namespace device
......
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ c8c016dd
......@@ -62,7 +62,7 @@ function(add_instance_library INSTANCE_NAME)
endforeach()
# Do not build mha instances if gfx94 or gfx90a targets are not on the target list
foreach(source IN LISTS ARGN)
if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx90a" AND source MATCHES "mha")
if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx90a" AND source MATCHES "mha")
message("removing mha instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
......@@ -87,6 +87,12 @@ function(add_instance_library INSTANCE_NAME)
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
foreach(source IN LISTS ARGN)
if(NOT INST_TARGETS MATCHES "gfx94" AND source MATCHES "gemm_xdl_universal_streamk" AND source MATCHES "_f8_")
message("removing gemm_universal_streamk_f8 instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
endif()
#only continue if there are some source files left on the list
if(ARGN)
......@@ -340,7 +346,7 @@ if(CK_DEVICE_CONV_INSTANCES)
endif()
if(CK_DEVICE_MHA_INSTANCES)
set(gpu_list ${INST_TARGETS})
if(gpu_list MATCHES "gfx94" OR gpu_list MATCHES "gfx90a")
if(gpu_list MATCHES "gfx94" OR gpu_list MATCHES "gfx90a")
add_library(device_mha_operations STATIC ${CK_DEVICE_MHA_INSTANCES})
add_library(composablekernels::device_mha_operations ALIAS device_mha_operations)
target_compile_features(device_mha_operations PUBLIC)
......
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