Skip to content

GitLab

"src/include/blockwise_gemm.hpp" did not exist on "66d5e5b344dd1fef553f7d430bda0a77ded5ffe7"
Commit 67ab3896 authored by Aleksander Dudek's avatar Aleksander Dudek
Browse files

Merge branch 'develop' into gemm_getname

parents 8adaf418 d5c8a334
Hide whitespace changes
Inline Side-by-side
Showing with 6039 additions and 111 deletions
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_b_scale.hpp 0 → 100644
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
namespace ck {
// Maximum Global Memory throughput pipeline with >=32KB data in fly
// GlobalPrefetchStages: >=2
// LocalPreFillStages: 1
// LocalPreFetchStages: 0
// LocalSharedMemoryBuffer: 1
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPacks>
struct BlockwiseGemmXdlops_pipeline_v2_b_scale
{
};
template <index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack
// ,bool TransposeC //disable transposec right now...
>
struct BlockwiseGemmXdlops_pipeline_v2_b_scale<BlockGemmPipelineScheduler::Intrawave,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
{
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>;
using Base::I0;
using Base::KRepeat;
using Base::xdlops_gemm;
using Base::CalculateCThreadOriginDataIndex;
using Base::CalculateCThreadOriginDataIndex8D;
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetCThreadBuffer;
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::a_block_desc_m0_m1_m2_k;
using Base::b_block_desc_n0_n1_n2_k;
using Base::AMmaKStride;
using Base::BMmaKStride;
static constexpr index_t WgpPerCU =
(4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
static constexpr index_t FullMemBandPrefetchStages = math::integer_divide_ceil(
32768 / WgpPerCU,
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);
static constexpr index_t PrefetchStages =
FullMemBandPrefetchStages >= 2
? FullMemBandPrefetchStages <= 8 ? FullMemBandPrefetchStages : 8
: 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = PrefetchStages;
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
if(num_loop % PrefetchStages == 1)
{
return TailNumber::One;
}
else if(num_loop % PrefetchStages == 2)
{
return TailNumber::Two;
}
else if(num_loop % PrefetchStages == 3)
{
return TailNumber::Three;
}
else if(num_loop % PrefetchStages == 4)
{
return TailNumber::Four;
}
else if(num_loop % PrefetchStages == 5)
{
return TailNumber::Five;
}
else if(num_loop % PrefetchStages == 6)
{
return TailNumber::Six;
}
else if(num_loop % PrefetchStages == 7)
{
return TailNumber::Seven;
}
else
{
return TailNumber::Full;
}
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
// Global prefetch [2, PrefetchStages]
static_for<1, PrefetchStages, 1>{}([&](auto iprefetch) {
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, iprefetch);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, iprefetch);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
});
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
static_for<0, PrefetchStages, 1>{}([&](auto iprefetch) {
// -------------------------------------------------------------------------------------------
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
block_sync_lds();
a_blockwise_copy.RunWrite(
a_block_desc, a_block_buf, Number<(iprefetch + 1) % PrefetchStages>{});
b_blockwise_copy.RunWrite(
b_block_desc, b_block_buf, Number<(iprefetch + 1) % PrefetchStages>{});
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, iprefetch);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, iprefetch);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
});
i += PrefetchStages;
} while(i < (num_loop - PrefetchStages));
}
// tail
auto LoopTailFunc = [&](auto tail_num) {
static_for<1, tail_num, 1>{}([&](auto iprefetch) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, iprefetch);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, iprefetch);
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
};
if constexpr(TailNum == TailNumber::One)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
}
else if constexpr(TailNum == TailNumber::Two)
{
LoopTailFunc(Number<2>{});
}
else if constexpr(TailNum == TailNumber::Three)
{
LoopTailFunc(Number<3>{});
}
else if constexpr(TailNum == TailNumber::Four)
{
LoopTailFunc(Number<4>{});
}
else if constexpr(TailNum == TailNumber::Five)
{
LoopTailFunc(Number<5>{});
}
else if constexpr(TailNum == TailNumber::Six)
{
LoopTailFunc(Number<6>{});
}
else if constexpr(TailNum == TailNumber::Seven)
{
LoopTailFunc(Number<7>{});
}
else if constexpr(TailNum == TailNumber::Full)
{
LoopTailFunc(Number<PrefetchStages>{});
}
}
protected:
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
using Base::b_thread_desc_;
using Base::c_thread_desc_;
};
template <index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack
// ,bool TransposeC //disable transposec right now...
>
struct BlockwiseGemmXdlops_pipeline_v2_b_scale<BlockGemmPipelineScheduler::Interwave,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
{
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>;
using Base::A_K1;
using Base::B_K1;
using Base::I0;
using Base::I1;
using Base::KPerThread;
using Base::xdlops_gemm;
using Base::CalculateCThreadOriginDataIndex;
using Base::CalculateCThreadOriginDataIndex8D;
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetCThreadBuffer;
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::a_block_desc_m0_m1_m2_k;
using Base::b_block_desc_n0_n1_n2_k;
static constexpr index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS;
static constexpr index_t KPerInnerLoop = math::max(KPerThread / NumMacClusters, KPack);
static constexpr index_t KRepeat = KPerThread / KPerInnerLoop;
static constexpr index_t WgpPerCU =
(4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
static constexpr index_t FullMemBandPrefetchStages = math::integer_divide_ceil(
32768 / WgpPerCU,
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);
static constexpr index_t PrefetchStages =
FullMemBandPrefetchStages >= 2
? FullMemBandPrefetchStages <= 8 ? FullMemBandPrefetchStages : 8
: 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = PrefetchStages;
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
if(num_loop % PrefetchStages == 1)
{
return TailNumber::One;
}
else if(num_loop % PrefetchStages == 2)
{
return TailNumber::Two;
}
else if(num_loop % PrefetchStages == 3)
{
return TailNumber::Three;
}
else if(num_loop % PrefetchStages == 4)
{
return TailNumber::Four;
}
else if(num_loop % PrefetchStages == 5)
{
return TailNumber::Five;
}
else if(num_loop % PrefetchStages == 6)
{
return TailNumber::Six;
}
else if(num_loop % PrefetchStages == 7)
{
return TailNumber::Seven;
}
else
{
return TailNumber::Full;
}
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename BScaleGridBuffer,
typename BScaleGridDesc,
typename BScaleThreadDesc,
typename BScaleThreadTransfer,
typename BScaleThreadTransferStep>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
const BScaleGridDesc& b_scale_grid_desc,
// BScaleThreadCopy
const BScaleThreadDesc& b_scale_thread_desc,
BScaleThreadTransfer& b_scale_thread_copy,
const BScaleGridBuffer& b_scale_grid_buf,
const BScaleThreadTransferStep& b_scale_thread_copy_step,
// num loop
index_t num_loop,
index_t num_loop_per_scale) const
{
ignore = num_loop_per_scale;
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_scale_thread_desc.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(n0, I0),
b_scale_thread_buf);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<0>{}));
});
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<1>{}));
// Initialize C
c_thread_buf.Clear();
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
// Global prefetch [2, PrefetchStages]
static_for<1, PrefetchStages, 1>{}([&](auto iprefetch) {
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, iprefetch);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, iprefetch);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
});
auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>(); // need?
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
static_for<0, PrefetchStages, 1>{}([&](auto iprefetch) {
// -------------------------------------------------------------------------------------------
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
// NOTE: Synchronize threads in a workgroup at the start of each MAC
// cluster, but except the first, as we can shorten non-MAC cluster a bit
// and there's no observable negative impact. The desired effect is waves in
// a workgroup executing MAC in sync. This avoids some out-of-sync waves
// hijacking MAC resource from other workgroups and reducing the chance of
// latency hiding by waiting for the rest of the workgroup at the eventual
// sync point.
if constexpr(k0.value != 0 || KRepeat == 1)
{
__builtin_amdgcn_s_barrier();
__builtin_amdgcn_sched_barrier(0);
}
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// The block_sync_lds() here performs double duty:
// A) safeguard against data hazard because barrier from
// blockwise_gemm is moved here B) reduce VMEM FIFO congestion
// by applying small delays to different wavefronts It is
// performed near the end of MAC cluster to minimize lgkmcnt
// penalty
if constexpr(k0.value == KRepeat - 1 &&
k_.value == KPerInnerLoop - KPack &&
m0.value == MRepeat - 1 && n0.value == NRepeat - 1)
{
__builtin_amdgcn_sched_barrier(0);
block_sync_lds();
__builtin_amdgcn_sched_barrier(0);
}
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
{
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(1);
__builtin_amdgcn_sched_barrier(0);
}
});
// static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t)
// {
// constexpr index_t c_offset =
// c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
// c_thread_buf(Number<c_offset>{}) +=
// c_thread_buf_per_scale[Number<t>{}] *
// type_convert<AccDataType>(b_scale_thread_buf[n0]);
// });
});
});
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(0);
__builtin_amdgcn_sched_barrier(0);
});
// static_for<0, NRepeat, 1>{}([&](auto n0) {
// b_scale_thread_copy.Run(b_scale_grid_desc,
// b_scale_grid_buf,
// b_scale_thread_desc,
// make_tuple(n0, I0),
// b_scale_thread_buf);
// b_scale_thread_copy.MoveSrcSliceWindow(
// b_scale_grid_desc, b_scale_thread_copy_step.At(Number<0>{}));
// });
// b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
// b_scale_thread_copy_step.At(Number<1>{}));
// block_sync_lds();
a_blockwise_copy.RunWrite(
a_block_desc, a_block_buf, Number<(iprefetch + 1) % PrefetchStages>{});
b_blockwise_copy.RunWrite(
b_block_desc, b_block_buf, Number<(iprefetch + 1) % PrefetchStages>{});
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, iprefetch);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, iprefetch);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
});
i += PrefetchStages;
} while(i < (num_loop - PrefetchStages));
}
// tail
auto LoopTailFunc = [&](auto tail_num) {
static_for<1, tail_num, 1>{}([&](auto iprefetch) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
if constexpr(k0.value != 0 || KRepeat == 1)
{
__builtin_amdgcn_s_barrier();
__builtin_amdgcn_sched_barrier(0);
}
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
if constexpr(k0.value == KRepeat - 1 &&
k_.value == KPerInnerLoop - KPack &&
m0.value == MRepeat - 1 && n0.value == NRepeat - 1)
{
__builtin_amdgcn_sched_barrier(0);
block_sync_lds();
__builtin_amdgcn_sched_barrier(0);
}
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
{
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(1);
__builtin_amdgcn_sched_barrier(0);
}
});
// static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
// constexpr index_t c_offset =
// c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
// c_thread_buf(Number<c_offset>{}) +=
// c_thread_buf_per_scale[Number<t>{}] *
// type_convert<AccDataType>(b_scale_thread_buf[n0]);
// });
});
});
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(0);
__builtin_amdgcn_sched_barrier(0);
});
// static_for<0, NRepeat, 1>{}([&](auto n0) {
// b_scale_thread_copy.Run(b_scale_grid_desc,
// b_scale_grid_buf,
// b_scale_thread_desc,
// make_tuple(n0, I0),
// b_scale_thread_buf);
// b_scale_thread_copy.MoveSrcSliceWindow(
// b_scale_grid_desc, b_scale_thread_copy_step.At(Number<0>{}));
// });
// b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
// b_scale_thread_copy_step.At(Number<1>{}));
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, iprefetch);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, iprefetch);
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
if constexpr(k0.value != 0 || KRepeat == 1)
{
__builtin_amdgcn_s_barrier();
__builtin_amdgcn_sched_barrier(0);
}
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
if constexpr(k0.value == KRepeat - 1 &&
k_.value == KPerInnerLoop - KPack &&
m0.value == MRepeat - 1 && n0.value == NRepeat - 1)
{
__builtin_amdgcn_sched_barrier(0);
block_sync_lds();
__builtin_amdgcn_sched_barrier(0);
}
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
{
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(1);
__builtin_amdgcn_sched_barrier(0);
}
});
// static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
// constexpr index_t c_offset =
// c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
// c_thread_buf(Number<c_offset>{}) +=
// c_thread_buf_per_scale[Number<t>{}] *
// type_convert<AccDataType>(b_scale_thread_buf[n0]);
// });
});
});
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(0);
__builtin_amdgcn_sched_barrier(0);
});
};
if constexpr(TailNum == TailNumber::One)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
if constexpr(k0.value != 0 || KRepeat == 1)
{
__builtin_amdgcn_s_barrier();
__builtin_amdgcn_sched_barrier(0);
}
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
if constexpr(k0.value == KRepeat - 1 &&
k_.value == KPerInnerLoop - KPack &&
m0.value == MRepeat - 1 && n0.value == NRepeat - 1)
{
__builtin_amdgcn_sched_barrier(0);
block_sync_lds();
__builtin_amdgcn_sched_barrier(0);
}
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
{
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(1);
__builtin_amdgcn_sched_barrier(0);
}
});
// static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
// constexpr index_t c_offset =
// c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
// c_thread_buf(Number<c_offset>{}) +=
// c_thread_buf_per_scale[Number<t>{}] *
// type_convert<AccDataType>(b_scale_thread_buf[n0]);
// });
});
});
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(0);
__builtin_amdgcn_sched_barrier(0);
});
}
else if constexpr(TailNum == TailNumber::Two)
{
LoopTailFunc(Number<2>{});
}
else if constexpr(TailNum == TailNumber::Three)
{
LoopTailFunc(Number<3>{});
}
else if constexpr(TailNum == TailNumber::Four)
{
LoopTailFunc(Number<4>{});
}
else if constexpr(TailNum == TailNumber::Five)
{
LoopTailFunc(Number<5>{});
}
else if constexpr(TailNum == TailNumber::Six)
{
LoopTailFunc(Number<6>{});
}
else if constexpr(TailNum == TailNumber::Seven)
{
LoopTailFunc(Number<7>{});
}
else if constexpr(TailNum == TailNumber::Full)
{
LoopTailFunc(Number<PrefetchStages>{});
}
}
protected:
// K->M loopover
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor(
make_tuple(Number<MRepeat>{}, I1, Number<KRepeat>{}, Number<KPerInnerLoop>{}),
make_tuple(Number<KPerInnerLoop>{},
Number<KRepeat * MRepeat * KPerInnerLoop>{},
Number<MRepeat * KPerInnerLoop>{},
I1));
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor(
make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPerInnerLoop>{}),
make_tuple(Number<KPerInnerLoop>{},
Number<KRepeat * NRepeat * KPerInnerLoop>{},
Number<NRepeat * KPerInnerLoop>{},
I1));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
ComputeDataType,
decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_),
Sequence<1, 1, 1, KPerInnerLoop>,
Sequence<0, 1, 2, 3>,
3,
A_K1,
A_K1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<BDataType,
ComputeDataType,
decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_),
Sequence<1, 1, 1, KPerInnerLoop>,
Sequence<0, 1, 2, 3>,
3,
B_K1,
B_K1>;
AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex()};
BThreadCopy b_thread_copy_{Base::CalculateBThreadOriginDataIndex()};
using Base::c_thread_desc_;
};
} // namespace ck
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_b_scale.hpp 0 → 100644
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
namespace ck {
// Compute optimized pipeline
// GlobalPrefetchStages: 2
// LocalPreFillStages: 1
// LocalPreFetchStages: 1
// LocalSharedMemoryBuffer: 1
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPacks>
struct BlockwiseGemmXdlops_pipeline_v3_b_scale
{
};
template <index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack
// ,bool TransposeC //disable transposec right now...
>
struct BlockwiseGemmXdlops_pipeline_v3_b_scale<BlockGemmPipelineScheduler::Intrawave,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
{
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>;
using Base::I0;
using Base::I1;
using Base::KRepeat;
using Base::xdlops_gemm;
using typename Base::HotLoopInstList;
using Base::CalculateCThreadOriginDataIndex;
using Base::CalculateCThreadOriginDataIndex8D;
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetCThreadBuffer;
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::a_block_desc_m0_m1_m2_k;
using Base::b_block_desc_n0_n1_n2_k;
using Base::AMmaKStride;
using Base::BMmaKStride;
static constexpr index_t PrefetchStages = 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 1;
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
ignore = num_loop;
return TailNumber::Full;
}
__device__ static constexpr auto HotLoopScheduler()
{
// A/B split schedule
// compiler is likely to use ds_read2 when instruction width smaller than 16bytes
constexpr auto num_ds_read_inst_a =
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16
? HotLoopInstList::A_LDS_Read_Inst_Num
: HotLoopInstList::A_LDS_Read_Inst_Num / 2;
constexpr auto num_ds_read_inst_b =
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16
? HotLoopInstList::B_LDS_Read_Inst_Num
: HotLoopInstList::B_LDS_Read_Inst_Num / 2;
constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
constexpr auto num_ds_write_inst_b = HotLoopInstList::B_LDS_Write_Inst_Num;
constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
constexpr auto ds_read_a_issue_cycle =
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
constexpr auto ds_read_b_issue_cycle =
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
constexpr auto ds_read_a_mfma_rate =
(mfma_cycle - 4 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
constexpr auto ds_read_b_mfma_rate =
(mfma_cycle - 4 + 2 * ds_read_b_issue_cycle - 1) / (2 * ds_read_b_issue_cycle);
constexpr auto num_dsread_a_mfma =
(num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
constexpr auto num_dsread_b_mfma =
(num_ds_read_inst_b + ds_read_b_mfma_rate - 1) / ds_read_b_mfma_rate;
// stage 1
// Separate this part?
// constexpr auto num_mfma_per_ds_read = sizeof(ComputeDataType) / sizeof(ADataType) >
// sizeof(ComputeDataType) / sizeof(BDataType)
// ? sizeof(ComputeDataType) / sizeof(ADataType)
// : sizeof(ComputeDataType) / sizeof(BDataType);
constexpr auto num_mfma_stage1 = num_mfma_inst - (num_dsread_a_mfma + num_dsread_b_mfma);
constexpr auto num_mfma_per_issue =
num_mfma_stage1 / (num_buffer_load_inst_a + num_buffer_load_inst_b);
constexpr auto num_dswrite_per_issue_a = num_ds_write_inst_a / num_buffer_load_inst_a;
constexpr auto num_dswrite_per_issue_b = num_ds_write_inst_b / num_buffer_load_inst_b;
static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) {
ignore = i;
static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) {
ignore = idswrite;
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA
});
static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
ignore = i;
static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) {
ignore = idswrite;
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_per_issue - num_dswrite_per_issue_b, 0); // MFMA
});
// stage 2
static_for<0, num_dsread_a_mfma, 1>{}([&](auto i) {
if constexpr((num_ds_read_inst_a - (i + 1) * ds_read_a_mfma_rate) >=
ds_read_a_mfma_rate)
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(0x100,
num_ds_read_inst_a - (num_dsread_a_mfma - 1) *
ds_read_a_mfma_rate,
0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
static_for<0, num_dsread_b_mfma, 1>{}([&](auto i) {
if constexpr((num_ds_read_inst_b - (i + 1) * ds_read_b_mfma_rate) >=
ds_read_b_mfma_rate)
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_b_mfma_rate, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(0x100,
num_ds_read_inst_b - (num_dsread_b_mfma - 1) *
ds_read_b_mfma_rate,
0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename BScaleGridBuffer,
typename BScaleGridDesc,
typename BScaleThreadDesc,
typename BScaleThreadTransfer,
typename BScaleThreadTransferStep>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
// BScaleThreadCopy
const BScaleGridDesc& b_scale_grid_desc,
const BScaleThreadDesc& b_scale_thread_desc,
BScaleThreadTransfer& b_scale_thread_copy,
const BScaleGridBuffer& b_scale_grid_buf,
const BScaleThreadTransferStep& b_scale_thread_copy_step,
// num loop
index_t num_loop,
index_t num_loop_per_scale) const
{
__builtin_amdgcn_sched_barrier(0);
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
// B scale buffer
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_scale_thread_desc.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(n0, I0),
b_scale_thread_buf);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<0>{}));
});
if(num_loop_per_scale == 1)
{
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<2>{}));
}
else
{
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<1>{}));
}
constexpr auto num_scale_k_block = BScaleThreadDesc{}.GetLength(I1);
constexpr auto num_scale_krepeat = KRepeat / num_scale_k_block;
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
// Global prefetch 2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// Local prefetch 1
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * BMmaKStride>{}),
b_block_buf,
b_scale_thread_buf[Number<n0 * num_scale_k_block + k0 / num_scale_krepeat>{}],
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(n0, I0),
b_scale_thread_buf);
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<0>{}));
});
if((i + 2) % num_loop_per_scale == 0)
{
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<2>{}));
}
else
{
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<1>{}));
}
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * BMmaKStride>{}),
b_block_buf,
b_scale_thread_buf[Number<n0 * num_scale_k_block +
k0 / num_scale_krepeat>{}],
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
i += 1;
} while(i < (num_loop - 1));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
__builtin_amdgcn_sched_barrier(0);
}
}
protected:
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
using Base::b_thread_desc_;
using Base::c_thread_desc_;
};
} // namespace ck
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4_b_scale.hpp 0 → 100644
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
namespace ck {
// Compute optimimal pipeline with highest resource request
// GlobalPrefetchStages: 4
// LocalPreFillStages: 2
// LocalPreFetchStages: 1
// LocalSharedMemoryBuffer: 2
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPacks>
struct BlockwiseGemmXdlops_pipeline_v4_b_scale
{
};
template <index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack
// ,bool TransposeC //disable transposec right now...
>
struct BlockwiseGemmXdlops_pipeline_v4_b_scale<BlockGemmPipelineScheduler::Intrawave,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
{
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>;
using Base::I0;
using Base::I1;
using Base::KRepeat;
using Base::xdlops_gemm;
using typename Base::HotLoopInstList;
using Base::CalculateCThreadOriginDataIndex;
using Base::CalculateCThreadOriginDataIndex8D;
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetCThreadBuffer;
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::a_block_desc_m0_m1_m2_k;
using Base::b_block_desc_n0_n1_n2_k;
using Base::AMmaKStride;
using Base::BMmaKStride;
static constexpr index_t PrefetchStages = 3;
static constexpr index_t PrefillStages = 2;
static constexpr index_t GlobalBufferNum = 1;
static constexpr index_t HotloopUnroll = 2;
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
if(num_loop % HotloopUnroll == 1)
{
return TailNumber::Odd;
}
else
{
return TailNumber::Even;
}
}
__device__ static constexpr void HotLoopScheduler()
{
// TODO: Take data type into consideration as pipe ver 3
// A-B splited schedule
constexpr auto num_ds_read_inst_a =
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16
? HotLoopInstList::A_LDS_Read_Inst_Num
: HotLoopInstList::A_LDS_Read_Inst_Num / 2;
constexpr auto num_ds_read_inst_b =
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16
? HotLoopInstList::B_LDS_Read_Inst_Num
: HotLoopInstList::B_LDS_Read_Inst_Num / 2;
constexpr auto num_issue_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
constexpr auto num_dswrite_per_issue_a =
(HotLoopInstList::A_LDS_Write_Inst_Num + num_issue_a - 1) / num_issue_a;
constexpr auto num_dsread_per_issue_a = num_ds_read_inst_a / num_issue_a;
constexpr auto num_issue_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
constexpr auto num_dswrite_per_issue_b =
(HotLoopInstList::B_LDS_Write_Inst_Num + num_issue_b - 1) / num_issue_b;
constexpr auto num_dsread_per_issue_b = num_ds_read_inst_b / num_issue_b;
constexpr auto num_mfma_per_issue =
HotLoopInstList::C_MFMA_Inst_Num / (num_issue_a + num_issue_b);
static_for<0, num_issue_a, 1>{}([&](auto i) {
ignore = i;
static_for<0, num_dsread_per_issue_a, 1>{}([&](auto idsread) {
ignore = idsread;
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) {
ignore = idswrite;
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(0x008,
num_mfma_per_issue - num_dsread_per_issue_a -
num_dswrite_per_issue_a,
0); // MFMA
});
static_for<0, num_issue_b, 1>{}([&](auto i) {
ignore = i;
static_for<0, num_dsread_per_issue_b, 1>{}([&](auto idsread) {
ignore = idsread;
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) {
ignore = idswrite;
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(0x008,
num_mfma_per_issue - num_dsread_per_issue_a -
num_dswrite_per_issue_b,
0); // MFMA
});
__builtin_amdgcn_sched_barrier(0);
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename BScaleGridBuffer,
typename BScaleGridDesc,
typename BScaleThreadDesc,
typename BScaleThreadTransfer,
typename BScaleThreadTransferStep>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
// BScaleThreadCopy
const BScaleGridDesc& b_scale_grid_desc,
const BScaleThreadDesc& b_scale_thread_desc,
BScaleThreadTransfer& b_scale_thread_copy,
const BScaleGridBuffer& b_scale_grid_buf,
const BScaleThreadTransferStep& b_scale_thread_copy_step,
// num loop
index_t num_loop,
index_t num_loop_per_scale) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
// B scale buffer
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_scale_thread_desc.GetElementSpaceSize());
StaticallyIndexedArray<decltype(a_thread_buf), Number<2>{}> a_thread_bufs;
StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
StaticallyIndexedArray<decltype(b_scale_thread_buf), Number<2>{}> b_scale_thread_bufs;
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(n0, I0),
b_scale_thread_bufs(I0));
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<0>{}));
});
if(num_loop_per_scale == 1)
{
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<2>{}));
}
else
{
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<1>{}));
}
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I0));
// Global prefetch 2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(n0, I0),
b_scale_thread_bufs(I1));
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<0>{}));
});
if(2 % num_loop_per_scale == 0)
{
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<2>{}));
}
else
{
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<1>{}));
}
// Local prefetch 1
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(I0));
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(I0),
b_scale_thread_bufs(I0)[n0],
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(I0));
});
});
});
// Local prefill 2
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I1));
// Global prefetch 3
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(n0, I0),
b_scale_thread_bufs(I0));
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<0>{}));
});
if(3 % num_loop_per_scale == 0)
{
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<2>{}));
}
else
{
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<1>{}));
}
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
// This hot loop has two legacy loopover, to implement the double local buffer strategy
do
{
auto LoopFunc = [&](auto lds_read_buf,
auto lds_read_reg_buf,
auto lds_write_buf,
auto mfma_reg_buf) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(lds_read_buf),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(lds_read_reg_buf));
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(lds_read_buf),
b_scale_thread_bufs(lds_read_buf)[n0],
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(lds_read_reg_buf));
});
});
// B scale copy
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(n0, I0),
b_scale_thread_bufs(lds_read_reg_buf));
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<0>{}));
});
if((i + 4 + mfma_reg_buf.value) % num_loop_per_scale == 0)
{
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<2>{}));
}
else
{
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<1>{}));
}
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(lds_write_buf));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(lds_write_buf));
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_bufs[mfma_reg_buf]
[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[mfma_reg_buf]
[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
HotLoopScheduler();
};
LoopFunc(I1, I1, I0, I0);
LoopFunc(I0, I0, I1, I1);
i += HotloopUnroll;
} while(i < (num_loop - PrefetchStages));
}
auto ReadWriteCompFunc = [&](auto lds_read_buf,
auto lds_read_reg_buf,
auto lds_write_buf,
auto mfma_reg_buf) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(lds_read_buf),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(lds_read_reg_buf));
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(lds_read_buf),
b_scale_thread_bufs(lds_read_buf)[n0],
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(lds_read_reg_buf));
});
});
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(lds_write_buf));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(lds_write_buf));
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
HotLoopScheduler();
};
auto ReadCompFunc = [&](auto lds_read_buf, auto lds_read_reg_buf, auto mfma_reg_buf) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(lds_read_buf),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(lds_read_reg_buf));
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(lds_read_buf),
b_scale_thread_bufs(lds_read_buf)[n0],
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(lds_read_reg_buf));
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
HotLoopScheduler();
};
auto CompFunc = [&](auto mfma_reg_buf) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
};
// tail
if constexpr(TailNum == TailNumber::Odd)
{
ReadWriteCompFunc(I1, I1, I0, I0);
ReadCompFunc(I0, I0, I1);
CompFunc(I0);
}
else if constexpr(TailNum == TailNumber::Even)
{
ReadCompFunc(I1, I1, I0);
CompFunc(I1);
}
}
protected:
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
using Base::b_thread_desc_;
using Base::c_thread_desc_;
};
} // namespace ck
include/ck/tensor_operation/gpu/device/device_gemm_v2.hpp
View file @ 67ab3896
...@@ -77,6 +77,43 @@ struct DeviceGemmV2R1 : public BaseOperator ...@@ -77,6 +77,43 @@ struct DeviceGemmV2R1 : public BaseOperator
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0; virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
}; };
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename BScaleType,
typename CDataType,
index_t ScaleBlockN,
index_t ScaleBlockK,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct DeviceGemmV2BScale : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t StrideA,
ck::index_t StrideB,
ck::index_t StrideC,
ck::index_t StrideScaleB,
const void* p_b_scale,
ck::index_t KSplit,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
virtual bool GetPermuteB() = 0;
virtual ck::index_t GetKPerBlock() = 0;
};
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp 100755 → 100644
View file @ 67ab3896
...@@ -469,7 +469,11 @@ struct DeviceGemm_Xdl_CShuffle_Streamk_V3 : public DeviceGemm_Streamk_V2<ALayout ...@@ -469,7 +469,11 @@ struct DeviceGemm_Xdl_CShuffle_Streamk_V3 : public DeviceGemm_Streamk_V2<ALayout
{ {
return false; return false;
} }
if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> &&
arg.Streamk_sel > 0)
{
return false;
}
if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding || if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::NKPadding || GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding || GemmSpec == GemmSpecialization::MNKPadding ||
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_b_scale.hpp 0 → 100644
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/host_utility/flush_cache.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_v2.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_b_scale.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename BScaleDataType,
typename CDataType,
typename GemmAccDataType,
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
GemmSpecialization GemmSpec,
index_t BlockSize,
index_t ScaleBlockN, // scale block for N
index_t ScaleBlockK, // scale block for K
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1,
index_t BK1,
index_t MPerXDL,
index_t NPerXDL,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
typename ComputeTypeA = CDataType,
typename ComputeTypeB = ComputeTypeA,
bool PermuteA = false,
bool PermuteB = false>
struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2BScale<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
BScaleDataType,
CDataType,
ScaleBlockN,
ScaleBlockK,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>
{
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_xdl_cshuffle_v3<
ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
GemmAccDataType,
CShuffleDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
GemmSpec,
BlockSize,
ScaleBlockN,
ScaleBlockK,
MPerBlock,
NPerBlock,
KPerBlock,
AK1,
BK1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
false,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
false,
BBlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CShuffleBlockTransferScalarPerVector_NPerBlock,
BlkGemmPipeSched,
BlkGemmPipelineVer,
ComputeTypeA,
ComputeTypeB,
PermuteA,
PermuteB>;
using Argument = typename GridwiseGemm::Argument;
// Invoker
struct Invoker : public BaseInvoker
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
if(stream_config.log_level_ > 0)
{
arg.Print();
}
if(!GridwiseGemm::CheckValidity(arg))
{
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
index_t gdx, gdy, gdz;
std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.KBatch);
float ave_time = 0;
index_t k_grain = arg.KBatch * KPerBlock;
index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
const auto Run = [&](const auto& kernel) {
if(stream_config.flush_cache)
{
Argument arg_ = arg;
const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
auto size_a_buffer =
a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType);
auto size_b_buffer =
b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType);
ck::utility::RotatingMemWrapper<Argument> rotating_mem(
arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer);
rotating_mem.Print();
auto run_flush_cache = [&]() {
// flush icache
ck::utility::flush_icache();
// rotating mem
rotating_mem.Next();
// clear c mem
if(arg_.KBatch > 1)
hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
0,
arg_.M * arg_.N * sizeof(CDataType),
stream_config.stream_id_));
};
ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
stream_config,
run_flush_cache,
kernel,
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
arg_);
}
else
{
if(arg.KBatch > 1)
hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
0,
arg.M * arg.N * sizeof(CDataType),
stream_config.stream_id_));
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
}
};
constexpr index_t minimum_occupancy =
BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave
? (BlkGemmPipelineVer == BlockGemmPipelineVersion::v3 &&
MPerBlock * NPerBlock * KPerBlock * sizeof(ADataType) <= 128 * 128 * 64 * 2)
? 2
: 1
: 2;
if(has_main_k_block_loop)
{
// Tail number always full
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1 ||
BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
{
if(arg.KBatch > 1)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy>;
Run(kernel);
}
}
// Tail number could be One to Seven
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
{
if(arg.KBatch > 1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::One>;
Run(kernel);
}
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Full)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Full>;
Run(kernel);
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Two>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Three)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Three>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Four)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Four>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Five)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Five>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Six>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Seven)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Seven>;
Run(kernel);
}
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::One>;
Run(kernel);
}
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Full)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Full>;
Run(kernel);
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Two>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Three)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Three>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Four)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Four>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Five)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Five>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Six>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Seven)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Seven>;
Run(kernel);
}
}
}
}
// Tail number could be Odd or Even
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
{
if(arg.KBatch > 1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds<
GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds<
GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_2lds<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_2lds<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
else
{
if(arg.KBatch > 1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
}
else
{
// Tail number always 1
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
{
if(arg.KBatch > 1)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
false,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy>;
Run(kernel);
}
}
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if(!ck::is_xdl_supported())
{
return false;
}
if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> && arg.KBatch > 1)
{
return false;
}
if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding ||
GemmSpec == GemmSpecialization::KPadding))
{
return false;
}
return GridwiseGemm::CheckValidity(arg);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
index_t GetKPerBlock() override { return KPerBlock; }
bool GetPermuteB() override { return PermuteB; }
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
CDataType* p_c,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t StrideScaleB,
const BScaleDataType* p_b_scale,
index_t KBatch,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
return Argument{p_a,
p_b,
p_c,
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideScaleB,
p_b_scale,
KBatch,
a_element_op,
b_element_op,
c_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t StrideScaleB,
const void* p_b_scale,
index_t KBatch,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideScaleB,
static_cast<const BScaleDataType*>(p_b_scale),
KBatch,
a_element_op,
b_element_op,
c_element_op);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
{BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
{BlockGemmPipelineScheduler::Interwave, "Interwave"}};
std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
{BlockGemmPipelineVersion::v1, "v1"},
{BlockGemmPipelineVersion::v2, "v2"},
{BlockGemmPipelineVersion::v3, "v3"},
{BlockGemmPipelineVersion::v4, "v4"},
{BlockGemmPipelineVersion::v5, "v5"}};
// clang-format off
str << "DeviceGemmXdlUniversal"
<< "<"
<< getGemmSpecializationString(GemmSpec) << ", "
<< std::string(ALayout::name)[0]
<< std::string(BLayout::name)[0]
<< std::string(CLayout::name)[0]
<< ">"
<< " BlkSize: "
<< BlockSize << ", "
<< "BlkTile: "
<< MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
<< "WaveTile: "
<< MPerXDL<<"x"<<NPerXDL << ", "
<< "WaveMap: "
<< MXdlPerWave<<"x" << NXdlPerWave<<", "
<< "VmemReadVec: "
<< ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
<< "BlkGemmPipelineScheduler: "
<< BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
<< "BlkGemmPipelineVersion: "
<< BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
<< "BlkGemmPipelinePrefetchStages: "
<< GridwiseGemm::BlockwiseGemmPipe::PrefetchStages;
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ 67ab3896
...@@ -44,6 +44,40 @@ __host__ __device__ inline half4_t pki4_to_half4(int q) ...@@ -44,6 +44,40 @@ __host__ __device__ inline half4_t pki4_to_half4(int q)
return res.template AsType<half4_t>()[Number<0>{}]; return res.template AsType<half4_t>()[Number<0>{}];
} }
__host__ __device__ inline half4_t pki4_to_half4_scale(int q, const ck::half2_t& scale)
{
const int LO = 0x000f000f;
const int HI = 0x00f000f0;
const int EX = 0x64006400;
// Extract the two int4 at low bit and create two fp16 number.
int lo = amd_assembly_and_or_b32(q, LO, EX);
// Extract the two int4 at hight bit and create two fp16 number.
int hi = amd_assembly_and_or_b32(q, HI, EX);
const int SUB = 0xE408E408; // half2 {-1032, -1032}
const int MUL = 0x2c002c00; // half2 {1 / 16, 1 / 16}
const int ADD = 0xd480d480; // half2 {-72, -72}
vector_type<half_t, 4> res;
res.template AsType<half2_t>()(Number<0>{}) =
amd_assembly_pk_add_f16(bit_cast<half2_t>(lo), bit_cast<half2_t>(SUB));
res.template AsType<half2_t>()(Number<1>{}) = amd_assembly_pk_fma_f16(
bit_cast<half2_t>(hi), bit_cast<half2_t>(MUL), bit_cast<half2_t>(ADD));
asm volatile("v_pk_mul_f16 %0, %1, %2"
: "=v"(res.template AsType<half2_t>()(Number<0>{}))
: "v"(res.template AsType<half2_t>()(Number<0>{})), "v"(scale));
asm volatile("v_pk_mul_f16 %0, %1, %2"
: "=v"(res.template AsType<half2_t>()(Number<1>{}))
: "v"(res.template AsType<half2_t>()(Number<1>{})), "v"(scale));
return res.template AsType<half4_t>()[Number<0>{}];
}
__host__ __device__ inline half2_t pki4_to_half2(pk_i4_t q) __host__ __device__ inline half2_t pki4_to_half2(pk_i4_t q)
{ {
#if 1 #if 1
...@@ -171,7 +205,42 @@ struct PassThroughPack8 ...@@ -171,7 +205,42 @@ struct PassThroughPack8
dst.template AsType<bhalf2_t>()(Number<3>{}) = dst.template AsType<bhalf2_t>()(Number<3>{}) =
pki4_to_bhalf2(src.template AsType<pk_i4_t>()[Number<3>{}]); pki4_to_bhalf2(src.template AsType<pk_i4_t>()[Number<3>{}]);
y = dst.template AsType<bhalf8_t>()[Number<0>{}]; y = dst.template AsType<bhalf8_t>()[Number<0>{}];
#endif
}
constexpr const static bool is_pack8_invocable = true;
};
struct DequantPack8
{
template <typename Y, typename X, typename Z>
__host__ __device__ void operator()(Y& y, const X& x, const Z& z) const;
__host__ __device__ constexpr void
operator()(ck::half8_t& y, const ck::pk_i4x4_t& x, const ck::half2_t& z) const
{
#if 1
vector_type<half_t, 8> result;
result.template AsType<half4_t>()(Number<0>{}) = pki4_to_half4_scale(bit_cast<int>(x), z);
result.template AsType<half4_t>()(Number<1>{}) =
pki4_to_half4_scale(bit_cast<int>(x) >> 8, z);
y = result.template AsType<half8_t>()[Number<0>{}];
#else
vector_type<half_t, 8> dst;
vector_type<pk_i4_t, 4> src{x};
dst.template AsType<half2_t>()(Number<0>{}) =
pki4_to_half2(src.template AsType<pk_i4_t>()[Number<0>{}]);
dst.template AsType<half2_t>()(Number<1>{}) =
pki4_to_half2(src.template AsType<pk_i4_t>()[Number<1>{}]);
dst.template AsType<half2_t>()(Number<2>{}) =
pki4_to_half2(src.template AsType<pk_i4_t>()[Number<2>{}]);
dst.template AsType<half2_t>()(Number<3>{}) =
pki4_to_half2(src.template AsType<pk_i4_t>()[Number<3>{}]);
y = dst.template AsType<half8_t>()[Number<0>{}];
#endif #endif
} }
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_b_scale.hpp 0 → 100644
View file @ 67ab3896
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_description/multi_index_transform_helper.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_scale_selector.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/utility/common_header.hpp"
namespace ck {
// Currently we do not have a elegant way to put single lds buffer & double lds buffer pipe in same
// kernel function Blockers:
// 1. Two separted declaration of __shared__ pointer is the key to make sure data access operate on
// two lds chunks.
// 2. Occupied __shared__ won't release until whole shader end, a.k.a AB and C may not use same lds
// buffer when we declare __shared__ inside blkgemmpipe
template <typename GridwiseGemm,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
index_t MinimumOccupancy = 1,
TailNumber TailNum = TailNumber::Full>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3(typename GridwiseGemm::Argument karg)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg);
GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
karg.p_c_grid + splitk_batch_offset.c_reduce_offset,
karg.p_b_scale_grid + splitk_batch_offset.scale_k_split_offset,
p_shared,
karg);
#else
ignore = karg;
#endif // end of if (defined(__gfx9__))
}
template <typename GridwiseGemm,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
index_t MinimumOccupancy = 1,
TailNumber TailNum = TailNumber::Full>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3_2lds(typename GridwiseGemm::Argument karg)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
// Pass two lds pointer is the key to tell compiler that ds_read/write
// operate on different lds chunk at same time without order dependecy
__shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];
__shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg);
GridwiseGemm::template Run_2Lds<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
karg.p_c_grid + splitk_batch_offset.c_reduce_offset,
karg.p_b_scale_grid + splitk_batch_offset.scale_k_split_offset,
p_shared_0,
p_shared_1,
karg);
#else
ignore = karg;
#endif // end of if (defined(__gfx9__))
}
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
tensor_operation::device::GemmSpecialization GemmSpec,
index_t BlockSize,
index_t ScaleBlockN, // scale N
index_t ScaleBlockK, // scale K
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1Value,
index_t BK1Value,
index_t MPerXdl,
index_t NPerXdl,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool AThreadTransferSrcResetCoordinateAfterRun,
index_t ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BThreadTransferSrcResetCoordinateAfterRun,
index_t BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v4,
typename ComputeTypeA = CDataType,
typename ComputeTypeB = ComputeTypeA,
bool PermuteA = false,
bool PermuteB = false>
struct GridwiseGemm_xdl_cshuffle_v3
{
using BScaleType = ck::half_t;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
static constexpr auto I6 = Number<6>{};
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
static constexpr auto AK0Number = Number<KPerBlock / AK1Value>{};
static constexpr auto BK0Number = Number<KPerBlock / BK1Value>{};
static constexpr auto AK1Number = Number<AK1Value>{};
static constexpr auto BK1Number = Number<BK1Value>{};
static constexpr index_t KPack =
math::max(math::lcm(AK1Number, BK1Number),
MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
static constexpr index_t APackedSize = []() {
if constexpr(is_same_v<remove_cvref_t<ADataType>, pk_i4_t>)
return 2;
else
return 1;
}();
static constexpr index_t BPackedSize = []() {
if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
return 2;
else
return 1;
}();
__host__ static auto CalculateGridSize(index_t M, index_t N, index_t KBatch)
{
return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, KBatch);
}
__host__ static auto CalculateMPadded(index_t M)
{
return math::integer_least_multiple(M, MPerBlock);
}
__host__ static auto CalculateNPadded(index_t N)
{
return math::integer_least_multiple(N, NPerBlock);
}
__host__ static auto CalculateKPadded(index_t K)
{
return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
}
__host__ static auto CalculateAK0Padded(index_t K, index_t K_Batch = 1)
{
auto K_t = K_Batch * KPerBlock;
return (K + K_t - 1) / K_t * (KPerBlock / AK1Value);
}
__host__ static auto CalculateBK0Padded(index_t K, index_t K_Batch = 1)
{
auto K_t = K_Batch * KPerBlock;
return (K + K_t - 1) / K_t * (KPerBlock / BK1Value);
}
__host__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
{
auto K_t = K_Batch * KPerBlock;
return (K + K_t - 1) / K_t * KPerBlock;
}
__host__ static auto CalculateKRead(index_t K, index_t K_Batch = 1)
{
constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
auto K_t = K_Batch * KReadVec;
return (K + K_t - 1) / K_t * KReadVec;
}
__host__ static auto CalculateMBlock(index_t M)
{
return math::integer_divide_ceil(M, MPerBlock);
}
__host__ static auto CalculateNBlock(index_t N)
{
return math::integer_divide_ceil(N, NPerBlock);
}
template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
__host__ __device__ static constexpr auto MakeGemmMmaTileDescriptor(const TileDesc_K0_MN_K1&)
{
constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
return transform_tensor_descriptor(
TileDesc_K0_MN_K1{},
make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
make_unmerge_transform(make_tuple(
Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
}
__host__ __device__ static auto MakeAGridDescriptor_AK0_M_AK1(
index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
{
const auto a_grid_desc_mraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
}();
using GemmSpecialization = tensor_operation::device::GemmSpecialization;
if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both M and K
const auto a_grid_desc_m_k =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
make_pass_through_transform(MPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad M, but not K
const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
make_right_pad_transform(M, MPad - M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad K, but not M
const auto a_grid_desc_m_k = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else
{
// not pad M or K
const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
}
__host__ __device__ static auto MakeBGridDescriptor_BK0_N_BK1(
index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
{
const auto b_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
}
}();
using GemmSpecialization = tensor_operation::device::GemmSpecialization;
static_assert(!(is_same_v<remove_cvref_t<ADataType>, pk_i4_t> &&
GemmSpec != GemmSpecialization::Default),
"pk_i4_t does not support padding");
if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both N and K
const auto b_grid_desc_n_k =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_right_pad_transform(N, NPad - N),
make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
make_pass_through_transform(NPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad N, but not K
const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad K, but not N
const auto b_grid_desc_n_k = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else
{
if constexpr(!PermuteB)
{
// not pad N or K
const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else
{
// Weight Tile Permute
constexpr index_t BK01 = KPerBlock / BK1Value;
// const index_t BK00 = BK0 / BK01;
const index_t BK0_ = StrideB / BK1Value;
const index_t BK00 = BK0_ / BK01;
const auto b_grid_desc_bk00_n_bk01_bk1_permute =
make_naive_tensor_descriptor_packed(make_tuple(BK00, N, BK01, BK1Value));
const auto b_grid_desc_bk0_n_bk1_permute = transform_tensor_descriptor(
b_grid_desc_bk00_n_bk01_bk1_permute,
make_tuple(make_merge_transform(make_tuple(BK00, BK01)),
make_pass_through_transform(make_tuple(N)),
make_pass_through_transform(BK1Value)),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return b_grid_desc_bk0_n_bk1_permute;
}
}
}
template <typename ABlockDesc_AK0_M_AK1>
__host__ __device__ static constexpr auto
MakeAMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
{
constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
return MakeGemmMmaTileDescriptor<MXdlPerWave, MWaves, MPerXdl>(ABlockDesc_AK0_M_AK1{});
}
template <typename BBlockDesc_BK0_N_BK1>
__host__ __device__ static constexpr auto
MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
{
constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
return MakeGemmMmaTileDescriptor<NXdlPerWave, NWaves, NPerXdl>(BBlockDesc_BK0_N_BK1{});
}
__host__ __device__ static auto
MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
{
const auto c_grid_desc_mraw_nraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
}
}();
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
#if 0
using GemmSpecialization = tensor_operation::device::GemmSpecialization;
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad M, but not N
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad N, but not M
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
// not pad M or N
return c_grid_desc_mraw_nraw;
}
#endif
}
struct Problem
{
__host__ Problem(index_t M_,
index_t N_,
index_t K_,
index_t StrideA_,
index_t StrideB_,
index_t StrideC_,
index_t StrideScaleB_,
index_t KBatch_)
: M{M_},
N{N_},
K{K_},
StrideA{StrideA_},
StrideB{StrideB_},
StrideC{StrideC_},
StrideScaleB{StrideScaleB_},
KBatch{KBatch_},
MPadded{CalculateMPadded(M_)},
NPadded{CalculateNPadded(N_)},
KRead{CalculateKRead(K_, KBatch_)},
KPadded{CalculateKPadded(K_, KBatch_)},
AK0{CalculateAK0Padded(K_, KBatch_)},
BK0{CalculateBK0Padded(K_, KBatch_)},
MBlock{CalculateMBlock(M_)},
NBlock{CalculateNBlock(N_)}
{
}
__host__ void Print() const
{
std::cout << "problem {"
<< "M:" << M << ", "
<< "N:" << N << ", "
<< "K:" << K << ", "
<< "SA:" << StrideA << ", "
<< "SB:" << StrideB << ", "
<< "SC:" << StrideC << ", "
<< "SScaleB:" << StrideScaleB << ", "
<< "MP:" << MPadded << ", "
<< "NP:" << NPadded << ", "
<< "KRead:" << KRead << ", "
<< "KP:" << KPadded << ", "
<< "AK0:" << AK0 << ", "
<< "BK0:" << BK0 << ", "
<< "MBlock: " << MBlock << ", "
<< "NBlock: " << NBlock << "}" << std::endl;
}
index_t M;
index_t N;
index_t K;
index_t StrideA;
index_t StrideB;
index_t StrideC;
index_t StrideScaleB;
index_t KBatch;
index_t MPadded;
index_t NPadded;
index_t KRead;
index_t KPadded;
index_t AK0;
index_t BK0;
index_t MBlock;
index_t NBlock;
};
// Argument
struct Argument : public tensor_operation::device::BaseArgument, public Problem
{
__host__ Argument(const ADataType* p_a_grid_,
const BDataType* p_b_grid_,
CDataType* p_c_grid_,
index_t M_,
index_t N_,
index_t K_,
index_t StrideA_,
index_t StrideB_,
index_t StrideC_,
index_t StrideScaleB_,
const BScaleType* p_b_scale_grid_,
index_t k_batch_,
AElementwiseOperation a_element_op_,
BElementwiseOperation b_element_op_,
CElementwiseOperation c_element_op_,
bool is_reduce_ = false)
: Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_, StrideScaleB_, k_batch_},
p_a_grid{p_a_grid_},
p_b_grid{p_b_grid_},
p_c_grid{p_c_grid_},
p_b_scale_grid{p_b_scale_grid_},
a_element_op{a_element_op_},
b_element_op{b_element_op_},
c_element_op{c_element_op_},
is_reduce(is_reduce_)
{
}
__host__ __device__ inline bool IsReduceAdd() const
{
return (Problem::KBatch > 1) && is_reduce;
}
__host__ __device__ inline bool IsAtomicAdd() const
{
return (Problem::KBatch > 1) && (!is_reduce);
}
const ADataType* p_a_grid;
const BDataType* p_b_grid;
CDataType* p_c_grid;
const BScaleType* p_b_scale_grid;
const AElementwiseOperation a_element_op;
const BElementwiseOperation b_element_op;
const CElementwiseOperation c_element_op;
bool is_reduce;
};
struct SplitKBatchOffset
{
__device__ SplitKBatchOffset(Argument& karg)
{
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
a_k_split_offset = blockIdx.z * karg.KRead / APackedSize;
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
a_k_split_offset = blockIdx.z * karg.KRead * karg.StrideA;
}
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
{
b_k_split_offset = blockIdx.z * karg.KRead * karg.StrideB;
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
{
if constexpr(!PermuteB)
{
b_k_split_offset = blockIdx.z * karg.KRead / BPackedSize;
}
else
{
const int k0_offset = karg.KRead * karg.N;
b_k_split_offset = blockIdx.z * k0_offset / BPackedSize;
}
}
// Calculate B scale offset
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
{
scale_k_split_offset = blockIdx.z * (karg.KRead / ScaleBlockK) * karg.StrideB;
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
{
scale_k_split_offset = blockIdx.z * (karg.KRead / ScaleBlockK);
}
if(blockIdx.z < static_cast<uint32_t>(karg.KBatch - 1))
{
karg.K = karg.KRead;
}
else
{
karg.K = karg.K - karg.KRead * (karg.KBatch - 1);
}
if(karg.IsReduceAdd())
{
c_reduce_offset = blockIdx.z * karg.M * karg.N;
}
else
{
c_reduce_offset = 0;
}
}
index_t a_k_split_offset;
index_t b_k_split_offset;
index_t scale_k_split_offset; // New member for scale matrix offset
index_t c_reduce_offset;
};
__device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
{
// A matrix in LDS memory, dst of blockwise copy
if constexpr(ABlockLdsExtraM || BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
{
return make_naive_tensor_descriptor(
make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
make_tuple(AK1Number, Number<KPerBlock + ABlockLdsExtraM>{}, I1));
}
// xor tensor transformation request more unnecessary vgpr usage, would cause register spill
// in some cases.
else if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
constexpr index_t LdsSize = 32 * 4 / KPerBlock / sizeof(ADataType) / APackedSize;
constexpr auto MLdsLayer = LdsSize < 1 ? 1 : LdsSize;
constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
make_tuple(
AK0Number * Number<MLdsLayer>{}, Number<MPerBlock / MLdsLayer>{}, AK1Number),
make_tuple(AK1Number, Number<KPerBlock * MLdsLayer>{}, I1));
constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
a_lds_block_desc,
make_tuple(make_xor_with_modulo_transform(make_tuple(
Number<MPerBlock / MLdsLayer>{}, Number<AK0Number * MLdsLayer>{})),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor(
a_lds_block_desc_permuted,
make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number<MLdsLayer>{})),
make_pass_through_transform(Number<MPerBlock / MLdsLayer>{}),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_lds_block_desc_ak0_mldslayer_m_ak1,
make_tuple(make_pass_through_transform(AK0Number),
make_merge_transform_v3_division_mod(
make_tuple(Number<MPerBlock / MLdsLayer>{}, Number<MLdsLayer>{})),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return a_lds_block_desc_ak0_m_ak1;
}
else // ColumnMajor A
{
// kfold and mpair dimension is not always required.
// more dimension in merge_transform increase the difficulty of generating immarg offset
// for compiler.
constexpr auto M0 = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1);
constexpr auto M1 = MPerBlock / M0;
constexpr auto KThreadWrite = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I0);
constexpr auto K0PerThreadWrite = AK0Number / KThreadWrite;
constexpr auto KThreadRead = 64 / MPerXdl;
constexpr auto K0PerThreadRead = AK0Number / KThreadRead;
constexpr auto kfold = (AK1Number * M0 * sizeof(ADataType) > 128)
? 1
: 128 / (AK1Number * M0 * sizeof(ADataType));
constexpr auto KThreadReadPerm =
(kfold * K0PerThreadWrite / K0PerThreadRead) > 1
? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
: KThreadRead;
// 1<=mpair<=n0
constexpr auto mpair = (AK1Number * MPerXdl * sizeof(ADataType) > 128)
? 1
: ((128 / (AK1Number * MPerXdl * sizeof(ADataType))) > M0
? M0
: 128 / (AK1Number * MPerXdl * sizeof(ADataType)));
constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<KThreadWrite / kfold / KThreadReadPerm>{},
Number<K0PerThreadWrite>{},
Number<KThreadReadPerm * M1>{},
Number<kfold * M0 / mpair>{},
Number<mpair>{},
AK1Number));
constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
a_lds_block_desc,
make_tuple(
make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(Number<K0PerThreadWrite>{}),
make_xor_with_modulo_transform(
make_tuple(Number<KThreadReadPerm * M1>{}, Number<kfold * M0 / mpair>{})),
make_pass_through_transform(Number<mpair>{}),
make_pass_through_transform(AK1Number)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}));
constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor(
a_lds_block_desc_permuted,
make_tuple(
make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(Number<K0PerThreadWrite>{}),
make_unmerge_transform(make_tuple(Number<KThreadReadPerm>{}, Number<M1>{})),
make_unmerge_transform(make_tuple(Number<kfold>{}, Number<M0 / mpair>{})),
make_pass_through_transform(Number<mpair>{}),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<0>{},
Sequence<1>{},
Sequence<2>{},
Sequence<3>{},
Sequence<4>{},
Sequence<5>{}),
make_tuple(Sequence<1>{},
Sequence<2>{},
Sequence<0, 3>{},
Sequence<4, 5>{},
Sequence<6>{},
Sequence<7>{}));
constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_lds_block_desc_unmerged,
make_tuple(make_merge_transform_v3_division_mod(
make_tuple(Number<KThreadReadPerm>{},
Number<KThreadWrite / kfold / KThreadReadPerm>{},
Number<kfold>{},
Number<K0PerThreadWrite>{})),
make_merge_transform_v3_division_mod(
make_tuple(Number<M0 / mpair>{}, Number<mpair>{}, Number<M1>{})),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return a_lds_block_desc_ak0_m_ak1;
}
}
__device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
{
// B matrix in LDS memory, dst of blockwise copy
if constexpr(BBlockLdsExtraN || BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
{
return make_naive_tensor_descriptor(
make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
make_tuple(BK1Number, Number<KPerBlock + BBlockLdsExtraN>{}, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
// NLdsLayer * K0 as logical Bank
constexpr index_t LdsSize = 32 * 4 / KPerBlock / sizeof(BDataType) / BPackedSize;
constexpr index_t NLdsLayer = LdsSize < 1 ? 1 : LdsSize;
constexpr auto b_lds_block_desc = make_naive_tensor_descriptor(
make_tuple(
BK0Number * Number<NLdsLayer>{}, Number<NPerBlock / NLdsLayer>{}, BK1Number),
make_tuple(BK1Number, Number<KPerBlock * NLdsLayer>{}, I1));
constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
b_lds_block_desc,
make_tuple(make_xor_with_modulo_transform(make_tuple(
Number<NPerBlock / NLdsLayer>{}, Number<BK0Number * NLdsLayer>{})),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
constexpr auto b_lds_block_desc_bk0_nldslayer_n_bk1 = transform_tensor_descriptor(
b_lds_block_desc_permuted,
make_tuple(make_unmerge_transform(make_tuple(BK0Number, Number<NLdsLayer>{})),
make_pass_through_transform(Number<NPerBlock / NLdsLayer>{}),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_lds_block_desc_bk0_nldslayer_n_bk1,
make_tuple(make_pass_through_transform(BK0Number),
make_merge_transform_v3_division_mod(
make_tuple(Number<NPerBlock / NLdsLayer>{}, Number<NLdsLayer>{})),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return b_lds_block_desc_bk0_n_bk1;
}
else // RowMajor B
{
constexpr auto N0 = BBlockTransferThreadClusterLengths_BK0_N_BK1{}.At(I1);
constexpr auto N1 = NPerBlock / N0;
constexpr auto KThreadWrite = BBlockTransferThreadClusterLengths_BK0_N_BK1{}.At(I0);
constexpr auto K0PerThreadWrite = BK0Number / KThreadWrite;
constexpr auto KThreadRead = 64 / NPerXdl;
constexpr auto K0PerThreadRead = BK0Number / KThreadRead;
constexpr auto kfold = (BK1Number * N0 * sizeof(BDataType) > 128)
? 1
: 128 / (BK1Number * N0 * sizeof(BDataType));
constexpr auto KThreadReadPerm =
(kfold * K0PerThreadWrite / K0PerThreadRead) > 1
? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
: KThreadRead;
// 1<=npair<=n0
constexpr auto npair = (BK1Number * NPerXdl * sizeof(BDataType) > 128)
? 1
: ((128 / (BK1Number * NPerXdl * sizeof(BDataType))) > N0
? N0
: 128 / (BK1Number * NPerXdl * sizeof(BDataType)));
constexpr auto b_lds_block_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<KThreadWrite / kfold / KThreadReadPerm>{},
Number<K0PerThreadWrite>{},
Number<KThreadReadPerm * N1>{},
Number<kfold * N0 / npair>{},
Number<npair>{},
BK1Number));
constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
b_lds_block_desc,
make_tuple(
make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(Number<K0PerThreadWrite>{}),
make_xor_with_modulo_transform(
make_tuple(Number<KThreadReadPerm * N1>{}, Number<kfold * N0 / npair>{})),
make_pass_through_transform(Number<npair>{}),
make_pass_through_transform(BK1Number)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}));
constexpr auto b_lds_block_desc_unmerged = transform_tensor_descriptor(
b_lds_block_desc_permuted,
make_tuple(
make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(Number<K0PerThreadWrite>{}),
make_unmerge_transform(make_tuple(Number<KThreadReadPerm>{}, Number<N1>{})),
make_unmerge_transform(make_tuple(Number<kfold>{}, Number<N0 / npair>{})),
make_pass_through_transform(Number<npair>{}),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<0>{},
Sequence<1>{},
Sequence<2>{},
Sequence<3>{},
Sequence<4>{},
Sequence<5>{}),
make_tuple(Sequence<1>{},
Sequence<2>{},
Sequence<0, 3>{},
Sequence<4, 5>{},
Sequence<6>{},
Sequence<7>{}));
constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_lds_block_desc_unmerged,
make_tuple(make_merge_transform_v3_division_mod(
make_tuple(Number<KThreadReadPerm>{},
Number<KThreadWrite / kfold / KThreadReadPerm>{},
Number<kfold>{},
Number<K0PerThreadWrite>{})),
make_merge_transform_v3_division_mod(
make_tuple(Number<N0 / npair>{}, Number<npair>{}, Number<N1>{})),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return b_lds_block_desc_bk0_n_bk1;
}
}
__device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
{
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
I1,
Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
}
using BlockwiseGemmPipe =
remove_cvref_t<decltype(BlockGemmPipeline_Selector<
BlkGemmPipelineVer,
BlkGemmPipeSched,
BlockSize,
ADataType,
BDataType,
ComputeTypeA,
AccDataType,
decltype(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()),
decltype(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()),
decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(
GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1())),
decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(
GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1())),
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXdl,
NPerXdl,
MXdlPerWave,
NXdlPerWave,
KPack>())>;
__device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
// LDS allocation for C shuffle in LDS
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
constexpr auto c_block_size =
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
return math::max((a_block_space_size_aligned * sizeof(ADataType) / APackedSize +
b_block_space_size_aligned * sizeof(BDataType) / BPackedSize),
c_block_size * sizeof(CShuffleDataType));
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
__host__ static constexpr bool CheckValidity(const Argument& karg)
{
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
"Invalid tuning param!");
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
!(is_same<tensor_layout::gemm::RowMajor, ALayout>::value))
{
if(!(karg.M % MPerBlock == 0))
{
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
std::cout << "Arg M value is not a multiple of MPerBlock! M: " << karg.M << " "
<< __FILE__ << ":" << __LINE__ << ", in function: " << __func__
<< std::endl;
}
return false;
}
}
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
(is_same<tensor_layout::gemm::RowMajor, BLayout>::value))
{
if(!(karg.N % NPerBlock == 0))
{
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
std::cout << "Arg N value is not a multiple of NPerBlock! N: " << karg.N << " "
<< __FILE__ << ":" << __LINE__ << ", in function: " << __func__
<< std::endl;
}
return false;
}
}
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
{
auto K_t = karg.KBatch * KPerBlock;
if(!(karg.K % K_t == 0))
{
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: "
<< karg.K << " " << __FILE__ << ":" << __LINE__
<< ", in function: " << __func__ << std::endl;
}
return false;
}
}
else
{
constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
auto K_t = karg.KBatch * KReadVec;
auto KReadPadSplited = math::integer_divide_ceil(karg.K, K_t) * KReadVec;
if((KReadPadSplited * (karg.KBatch - 1)) >= karg.K)
{
return false;
}
}
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
if(karg.K % ABlockTransferSrcScalarPerVector != 0)
{
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
std::cout << "Arg K (" << karg.K
<< ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
<< ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
<< __LINE__ << ", in function: " << __func__ << std::endl;
}
return false;
}
}
else
{
if(karg.M % ABlockTransferSrcScalarPerVector != 0)
{
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
std::cout << "Arg M (" << karg.M
<< ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
<< ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
<< __LINE__ << ", in function: " << __func__ << std::endl;
}
return false;
}
}
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
if(karg.N % BBlockTransferSrcScalarPerVector != 0)
{
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
std::cout << "Arg N (" << karg.N
<< ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
<< BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
<< __LINE__ << ", in function: " << __func__ << std::endl;
}
return false;
}
}
else
{
if(karg.K % BBlockTransferSrcScalarPerVector != 0)
{
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
std::cout << "Arg K (" << karg.K
<< ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
<< BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
<< __LINE__ << ", in function: " << __func__ << std::endl;
}
return false;
}
}
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
if(karg.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
{
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
std::cout << "Arg N (" << karg.N
<< ") value is not a multiple of "
"CShuffleBlockTransferScalarPerVector_NPerBlock ("
<< CShuffleBlockTransferScalarPerVector_NPerBlock << " )! "
<< __FILE__ << ":" << __LINE__ << ", in function: " << __func__
<< std::endl;
}
return false;
}
}
else
{
if(karg.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
{
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
std::cout << "Arg M (" << karg.M
<< ") value is not a multiple of "
"CShuffleBlockTransferScalarPerVector_NPerBlock ("
<< CShuffleBlockTransferScalarPerVector_NPerBlock << " )! "
<< __FILE__ << ":" << __LINE__ << ", in function: " << __func__
<< std::endl;
}
return false;
}
}
if constexpr(!(is_same<remove_cvref_t<CDataType>, half_t>::value ||
is_same<remove_cvref_t<CDataType>, float>::value ||
is_same<remove_cvref_t<CDataType>, bhalf_t>::value ||
is_same<remove_cvref_t<CDataType>, int32_t>::value))
{
if(!karg.IsReduceAdd())
{
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
std::cout << " KBatch: " << karg.KBatch << " > 1 is not support yet" << __FILE__
<< ":" << __LINE__ << ", in function: " << __func__ << std::endl;
}
if(karg.KBatch > 1)
{
return false;
}
}
}
// check gridwise gemm pipeline
const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value);
if constexpr(BlkGemmPipelineVer != BlockGemmPipelineVersion::v1)
{
if(num_k_loop <= BlockwiseGemmPipe::PrefetchStages)
{
return false;
}
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return true;
}
__host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
{
const index_t num_loop = K / KPerBlock;
return BlockwiseGemmPipe::BlockHasHotloop(num_loop);
}
__host__ static constexpr TailNumber CalculateKBlockLoopTailNum(index_t K)
{
const index_t num_loop = K / KPerBlock;
return BlockwiseGemmPipe::BlockLoopTailNum(num_loop);
}
template <typename CGridDesc>
__host__ __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock)
{
const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
return c_grid_desc_mblock_mperblock_nblock_nperblock;
}
// return block_id to C matrix tile idx (m0, n0) mapping
// if arch = gfx942
using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
// using Block2CTileMap = BlockToCTileMap_3DGrid_KSplit<MPerBlock, NPerBlock>;
template <typename AGridDesc_AK0_M_K1,
typename BGridDesc_BK0_N_K1,
typename BScaleGridDesc_BN_AK,
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd>
__device__ static void Run(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
const BScaleType* p_b_scale_grid,
void* p_shared,
const Problem& problem,
const AGridDesc_AK0_M_K1& a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_K1& b_grid_desc_bk0_n_bk1,
const BScaleGridDesc_BN_AK& b_scale_grid_desc_bn_ak,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
// B Scale buffer
const auto b_scale_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_scale_grid, b_scale_grid_desc_bn_ak.GetElementSpaceSize());
const AElementwiseOperation a_element_op{};
const BElementwiseOperation b_element_op{};
const CElementwiseOperation c_element_op{};
// divide block work by [M, N]
const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4};
const auto block_work_idx =
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_ctile_map.ValidCTileIndex(
block_work_idx,
make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<AK0Number, MPerBlock, AK1Number>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ADataType,
ADataType,
decltype(a_grid_desc_ak0_m_ak1),
decltype(a_block_desc_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
ABlockTransferSrcVectorDim,
2,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true,
BlockwiseGemmPipe::GlobalBufferNum>(
a_grid_desc_ak0_m_ak1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_ak0_m_ak1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<BK0Number, NPerBlock, BK1Number>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BDataType,
BDataType,
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_block_desc_bk0_n_bk1),
BBlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
BBlockTransferSrcVectorDim,
2,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true,
BlockwiseGemmPipe::GlobalBufferNum>(
b_grid_desc_bk0_n_bk1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
// Cast after lds
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ADataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
reinterpret_cast<BDataType*>(static_cast<char*>(p_shared) + a_block_space_size_aligned *
sizeof(ADataType) /
APackedSize),
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1Number, 0, 0);
// Blockwise GEMM pipeline
static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
// b scale
// static_assert(KPerBlock <= ScaleBlockK);
static constexpr auto mfma = MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl>{};
static constexpr auto KPerXdlops = mfma.GetKPerXdlops();
static constexpr auto K1PerXdlops = mfma.GetK1PerXdlops();
static constexpr auto K0PerXdlops = KPerXdlops / K1PerXdlops;
static constexpr auto KPerThread = KPerBlock / K0PerXdlops;
static constexpr auto ScaleSliceSizeN = NXdlPerWave;
static constexpr auto ScaleSliceSizeK = (KPerThread + ScaleBlockK - 1) / ScaleBlockK;
static constexpr auto KBlockScaleSliceSizeK = (KPerBlock + ScaleBlockK - 1) / ScaleBlockK;
constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<ScaleSliceSizeN>{}, Number<ScaleSliceSizeK>{}));
constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
auto b_thread_offset_n =
get_thread_local_1d_id() % NPerXdl + (get_thread_local_1d_id() / 64) % NWaves * NPerXdl;
auto b_thread_offset_k = (get_thread_local_1d_id() % 64) / NPerXdl * KPerThread;
auto b_scale_thread_copy =
ThreadwiseTensorSliceTransfer_v2<BScaleType,
BScaleType,
decltype(b_scale_grid_desc_bn_ak),
decltype(b_scale_thread_desc),
Sequence<1, ScaleSliceSizeK>,
Sequence<0, 1>,
1,
ScaleSliceSizeK,
1,
false>(
b_scale_grid_desc_bn_ak,
make_multi_index(block_n_id * NPerBlock / ScaleBlockN + b_thread_offset_n,
b_thread_offset_k / ScaleBlockK));
constexpr auto b_scale_thread_slice_copy_step =
make_tuple(make_multi_index(NWaves * NPerXdl, 0),
make_multi_index(-NPerBlock, 0),
make_multi_index(-NPerBlock, KBlockScaleSliceSizeK));
const index_t num_k_block_per_scale = (ScaleBlockK + KPerBlock - 1) / KPerBlock;
blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(
a_grid_desc_ak0_m_ak1,
a_block_desc_ak0_m_ak1,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_grid_desc_bk0_n_bk1,
b_block_desc_bk0_n_bk1,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
c_thread_buf,
b_scale_grid_desc_bn_ak,
b_scale_thread_desc,
b_scale_thread_copy,
b_scale_grid_buf,
b_scale_thread_slice_copy_step,
num_k_block_main_loop,
num_k_block_per_scale);
// shuffle C and write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<CShuffleDataType*>(p_shared),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
CShuffleDataType,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
// shuffle: blockwise copy C from LDS to global
auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
ThisThreadBlock, // ThreadGroup
CElementwiseOperation, // ElementwiseOperation,
CGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
CShuffleDataType, // typename SrcData,
CDataType, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
true, // bool ThreadTransferSrcResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_m_id, 0, block_n_id, 0),
c_element_op};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
M2,
1,
M4,
1>>{};
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_c_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
c_shuffle_block_copy_lds_to_global.Run(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
c_shuffle_block_buf,
c_grid_desc_mblock_mperblock_nblock_nperblock,
c_grid_buf);
if constexpr(access_id < num_access - 1)
{
constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
// move on C
c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
}
});
}
}
template <bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd>
__device__ static void Run(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
const BScaleType* p_b_scale_grid,
void* p_shared,
const Problem& problem)
{
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n, problem.MBlock, problem.NBlock);
// B Scale grid
const auto b_scale_grid_desc_bn_ak = make_naive_tensor_descriptor(
make_tuple(math::integer_divide_ceil(problem.N, ScaleBlockN),
math::integer_divide_ceil(problem.K, ScaleBlockK)),
make_tuple(problem.StrideScaleB, 1));
Run<decltype(a_grid_desc_ak0_m_ak1),
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_scale_grid_desc_bn_ak),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
HasMainKBlockLoop,
CGlobalMemoryDataOperation,
TailNum>(p_a_grid,
p_b_grid,
p_c_grid,
p_b_scale_grid,
p_shared,
problem,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
b_scale_grid_desc_bn_ak,
c_grid_desc_mblock_mperblock_nblock_nperblock);
}
template <typename AGridDesc_AK0_M_K1,
typename BGridDesc_BK0_N_K1,
typename BScaleGridDesc_BN_AK,
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd>
__device__ static void Run_2Lds(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
const BScaleType* p_b_scale_grid,
void* p_shared_0,
void* p_shared_1,
const Problem& problem,
const AGridDesc_AK0_M_K1& a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_K1& b_grid_desc_bk0_n_bk1,
const BScaleGridDesc_BN_AK& b_scale_grid_desc_bn_ak,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
// B Scale buffer
const auto b_scale_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_scale_grid, b_scale_grid_desc_bn_ak.GetElementSpaceSize());
const AElementwiseOperation a_element_op{};
const BElementwiseOperation b_element_op{};
const CElementwiseOperation c_element_op{};
// divide block work by [M, N]
const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4};
const auto block_work_idx =
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_ctile_map.ValidCTileIndex(
block_work_idx,
make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<AK0Number, MPerBlock, AK1Number>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ADataType,
ADataType,
decltype(a_grid_desc_ak0_m_ak1),
decltype(a_block_desc_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
ABlockTransferSrcVectorDim,
2,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true,
BlockwiseGemmPipe::GlobalBufferNum>(
a_grid_desc_ak0_m_ak1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_ak0_m_ak1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<BK0Number, NPerBlock, BK1Number>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BDataType,
BDataType,
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_block_desc_bk0_n_bk1),
BBlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
BBlockTransferSrcVectorDim,
2,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true,
BlockwiseGemmPipe::GlobalBufferNum>(
b_grid_desc_bk0_n_bk1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ADataType*>(p_shared_0), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
bit_cast<BDataType*>(static_cast<char*>(p_shared_0) +
a_block_space_size_aligned * sizeof(ADataType) / APackedSize),
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
auto a_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ADataType*>(p_shared_1), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
bit_cast<BDataType*>(bit_cast<char*>(p_shared_1) +
a_block_space_size_aligned * sizeof(ADataType) / APackedSize),
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
auto a_block_bufs = make_tuple(a_block_buf_ping, a_block_buf_pong);
auto b_block_bufs = make_tuple(b_block_buf_ping, b_block_buf_pong);
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1Number, 0, 0);
// Blockwise GEMM pipeline
static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
// B scale
static constexpr auto mfma = MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl>{};
static constexpr auto KPerXdlops = mfma.GetKPerXdlops();
static constexpr auto K1PerXdlops = mfma.GetK1PerXdlops();
static constexpr auto K0PerXdlops = KPerXdlops / K1PerXdlops;
static constexpr auto KPerThread = KPerBlock / K0PerXdlops;
const index_t ScaleSliceSizeN = NXdlPerWave;
static constexpr auto ScaleSliceSizeK = (KPerThread + ScaleBlockK - 1) / ScaleBlockK;
static constexpr auto KBlockScaleSliceSizeK = (KPerBlock + ScaleBlockK - 1) / ScaleBlockK;
constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<ScaleSliceSizeN>{}, Number<ScaleSliceSizeK>{}));
constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
auto b_thread_offset_n =
get_thread_local_1d_id() % NPerXdl + (get_thread_local_1d_id() / 64) % NWaves * NPerXdl;
auto b_thread_offset_k = (get_thread_local_1d_id() % 64) / NPerXdl * KPerThread;
auto b_scale_thread_copy =
ThreadwiseTensorSliceTransfer_v2<BScaleType,
BScaleType,
decltype(b_scale_grid_desc_bn_ak),
decltype(b_scale_thread_desc),
Sequence<1, ScaleSliceSizeK>,
Sequence<0, 1>,
1,
ScaleSliceSizeK,
1,
false>(
b_scale_grid_desc_bn_ak,
make_multi_index(block_n_id * NPerBlock / ScaleBlockN + b_thread_offset_n,
b_thread_offset_k / ScaleBlockK));
constexpr auto b_scale_thread_slice_copy_step =
make_tuple(make_multi_index(NWaves * NPerXdl, 0),
make_multi_index(-NPerBlock, 0),
make_multi_index(-NPerBlock, KBlockScaleSliceSizeK));
const index_t num_k_block_per_scale = (ScaleBlockK + KPerBlock - 1) / KPerBlock;
blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(
a_grid_desc_ak0_m_ak1,
a_block_desc_ak0_m_ak1,
a_blockwise_copy,
a_grid_buf,
a_block_bufs,
a_block_slice_copy_step,
b_grid_desc_bk0_n_bk1,
b_block_desc_bk0_n_bk1,
b_blockwise_copy,
b_grid_buf,
b_block_bufs,
b_block_slice_copy_step,
c_thread_buf,
b_scale_grid_desc_bn_ak,
b_scale_thread_desc,
b_scale_thread_copy,
b_scale_grid_buf,
b_scale_thread_slice_copy_step,
num_k_block_main_loop,
num_k_block_per_scale);
// shuffle C and write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<CShuffleDataType*>(p_shared_0),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
CShuffleDataType,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
// shuffle: blockwise copy C from LDS to global
auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
ThisThreadBlock, // ThreadGroup
CElementwiseOperation, // ElementwiseOperation,
CGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
CShuffleDataType, // typename SrcData,
CDataType, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
true, // bool ThreadTransferSrcResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_m_id, 0, block_n_id, 0),
c_element_op};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
M2,
1,
M4,
1>>{};
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_c_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
c_shuffle_block_copy_lds_to_global.Run(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
c_shuffle_block_buf,
c_grid_desc_mblock_mperblock_nblock_nperblock,
c_grid_buf);
if constexpr(access_id < num_access - 1)
{
constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
// move on C
c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
}
});
}
}
template <bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd>
__device__ static void Run_2Lds(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
const BScaleType* p_b_scale_grid,
void* p_shared_0,
void* p_shared_1,
const Problem& problem)
{
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n, problem.MBlock, problem.NBlock);
const auto b_scale_grid_desc_bn_ak = make_naive_tensor_descriptor(
make_tuple(math::integer_divide_ceil(problem.N, ScaleBlockN),
math::integer_divide_ceil(problem.K, ScaleBlockK)),
make_tuple(problem.StrideScaleB, 1));
Run_2Lds<decltype(a_grid_desc_ak0_m_ak1),
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_scale_grid_desc_bn_ak),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
HasMainKBlockLoop,
CGlobalMemoryDataOperation,
TailNum>(p_a_grid,
p_b_grid,
p_c_grid,
p_b_scale_grid,
p_shared_0,
p_shared_1,
problem,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
b_scale_grid_desc_bn_ak,
c_grid_desc_mblock_mperblock_nblock_nperblock);
}
};
} // namespace ck
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
View file @ 67ab3896
...@@ -1222,6 +1222,206 @@ struct ThreadwiseTensorSliceTransfer_v4 ...@@ -1222,6 +1222,206 @@ struct ThreadwiseTensorSliceTransfer_v4
}); });
} }
// Fuse scale
template <typename SrcRefToOriginDisplacement,
typename DstOriginIdx,
typename SrcBuffer,
typename DstBuffer>
__device__ void Run(const SrcDesc&,
const SrcRefToOriginDisplacement&,
const SrcBuffer& src_buf,
const DstData& scale,
const DstDesc&,
const DstOriginIdx&,
DstBuffer& dst_buf) const
{
static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
"wrong! SrcDesc and DstDesc need to known at compile-time");
static_assert(
is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value &&
is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
"wrong! SrcBuffer or DstBuffer data type is wrong");
static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
static_assert(is_known_at_compile_time<remove_cvref_t<SrcRefToOriginDisplacement>>::value &&
is_known_at_compile_time<remove_cvref_t<DstOriginIdx>>::value,
"wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known "
"at compile-time");
// SrcDesc and DstDesc are known at compile-time
constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
// SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time
constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{});
constexpr auto dst_origin_idx = to_multi_index(DstOriginIdx{});
// scalar per access of each dim
constexpr auto src_scalar_per_access = generate_sequence_v2(
[&](auto i) constexpr {
if constexpr(i == SrcVectorDim)
{
return Number<SrcScalarPerVector>{};
}
else
{
return Number<1>{};
}
},
Number<nDim>{});
// scalar step (if steping on SrcVectorDim) of each dim
constexpr auto src_scalar_step_in_vector = generate_sequence_v2(
[&](auto i) constexpr {
if constexpr(i == SrcVectorDim)
{
return Number<1>{};
}
else
{
return Number<0>{};
}
},
Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / src_scalar_per_access;
constexpr auto dim_access_order = DimAccessOrder{};
constexpr auto ordered_access_lengths =
container_reorder_given_new2old(access_lengths, dim_access_order);
static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
#if 0
// TODO: unable to compile
// position in slice window
constexpr auto data_to_origin_disp_idx =
container_reorder_given_old2new(ordered_access_idx, dim_access_order) *
src_scalar_per_access;
#else
// position in slice window
constexpr auto data_to_origin_disp_idx =
ordered_access_idx.ReorderGivenOld2New(dim_access_order) * src_scalar_per_access;
#endif
// src coordinate
constexpr auto src_ref_to_data_disp_idx =
src_ref_to_origin_disp_idx + data_to_origin_disp_idx;
constexpr auto src_ref_to_data_disp_coord_step =
make_tensor_coordinate_step(src_desc, src_ref_to_data_disp_idx);
auto src_data_coord = src_ref_coord_;
move_tensor_coordinate(src_desc, src_data_coord, src_ref_to_data_disp_coord_step);
vector_type_maker_t<SrcData, SrcScalarPerVector / PackedSize> src_tmp_vector;
using src_vector_t = typename decltype(src_tmp_vector)::type;
const bool is_src_valid = coordinate_has_valid_offset_assuming_visible_index_is_valid(
src_desc, src_data_coord);
// copy data from src_buf into src_tmp_vector
if constexpr(SrcBuffer::IsDynamicBuffer())
{
src_tmp_vector.template AsType<src_vector_t>()(Number<0>{}) =
src_buf.template Get<src_vector_t>(src_data_coord.GetOffset() / PackedSize,
is_src_valid);
}
else if constexpr(SrcBuffer::IsStaticBuffer())
{
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
constexpr index_t src_offset = src_desc.CalculateOffset(
src_ref_to_origin_disp_idx + data_to_origin_disp_idx +
i * src_scalar_step_in_vector);
src_tmp_vector.template AsType<SrcData>()(i) = src_buf[Number<src_offset>{}];
});
}
if constexpr(is_same<remove_cvref_t<SrcData>, pk_i4_t>::value)
{
// copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
// DstData)
vector_type_maker_t<DstData, SrcScalarPerVector> dst_tmp_vector;
vector_type<DstData, 2> scale_vector;
scale_vector.template AsType<DstData>()(Number<0>{}) = scale;
scale_vector.template AsType<DstData>()(Number<1>{}) = scale;
constexpr index_t pack_size = 8;
static_assert(SrcScalarPerVector % pack_size == 0, "");
using src_v_t = typename vector_type_maker_t<SrcData, pack_size / PackedSize>::type;
using dst_v_t = typename vector_type_maker_t<DstData, pack_size>::type;
using scale_v_t = typename vector_type_maker_t<DstData, 2>::type;
static_for<0, SrcScalarPerVector / pack_size, 1>{}([&](auto i) {
ck::tensor_operation::element_wise::DequantPack8{}(
dst_tmp_vector.template AsType<dst_v_t>()(i),
src_tmp_vector.template AsType<src_v_t>()[i],
scale_vector.template AsType<scale_v_t>()[Number<0>{}]);
});
// copy data from dst_tmp_vector into dst_buf
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
constexpr index_t dst_offset = dst_desc.CalculateOffset(
dst_origin_idx + data_to_origin_disp_idx + i * src_scalar_step_in_vector);
dst_buf(Number<dst_offset>{}) = dst_tmp_vector.template AsType<DstData>()[i];
});
}
else if constexpr(is_same<remove_cvref_t<SrcData>, f8_t>::value &&
is_same<remove_cvref_t<DstData>, half_t>::value &&
SrcScalarPerVector % 2 == 0)
{
// copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
// DstData)
vector_type_maker_t<DstData, SrcScalarPerVector> dst_tmp_vector;
constexpr index_t pack_size = 2;
using dst_v_t = typename vector_type_maker_t<DstData, pack_size>::type;
using src_v_t = typename vector_type_maker_t<SrcData, pack_size>::type;
static_for<0, SrcScalarPerVector / pack_size, 1>{}([&](auto i) {
ck::tensor_operation::element_wise::PassThroughPack2{}(
dst_tmp_vector.template AsType<dst_v_t>()(i),
src_tmp_vector.template AsType<src_v_t>()[i]);
});
// copy data from dst_tmp_vector into dst_buf
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
constexpr index_t dst_offset = dst_desc.CalculateOffset(
dst_origin_idx + data_to_origin_disp_idx + i * src_scalar_step_in_vector);
dst_buf(Number<dst_offset>{}) = dst_tmp_vector.template AsType<DstData>()[i];
});
}
else
{
// copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
// DstData)
vector_type_maker_t<DstData, SrcScalarPerVector> dst_tmp_vector;
// TODO: if SrcData and DstData are vetor type, then static_cast may not compile
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
dst_tmp_vector.template AsType<DstData>()(i) =
type_convert<DstData>(src_tmp_vector.template AsType<SrcData>()[i]);
});
// copy data from dst_tmp_vector into dst_buf
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
constexpr index_t dst_offset = dst_desc.CalculateOffset(
dst_origin_idx + data_to_origin_disp_idx + i * src_scalar_step_in_vector);
dst_buf(Number<dst_offset>{}) = dst_tmp_vector.template AsType<DstData>()[i];
});
}
});
}
template <typename SrcSliceMoveStepIdx> template <typename SrcSliceMoveStepIdx>
__device__ void MoveSrcSliceWindow(const SrcDesc&, __device__ void MoveSrcSliceWindow(const SrcDesc&,
const SrcSliceMoveStepIdx& src_slice_move_step_idx) const SrcSliceMoveStepIdx& src_slice_move_step_idx)
...@@ -1344,7 +1544,7 @@ struct ThreadwiseTensorSliceTransfer_StaticToStatic ...@@ -1344,7 +1544,7 @@ struct ThreadwiseTensorSliceTransfer_StaticToStatic
ElementwiseOperation element_op_; ElementwiseOperation element_op_;
}; };
// Specilized for WMMA-Navi3 // Specialized for gfx11
// A single Wave32 is composed by double row // A single Wave32 is composed by double row
// Data exchange allowed between these two rows // Data exchange allowed between these two rows
// This RowLane Dst buf will be filled from two Src buf // This RowLane Dst buf will be filled from two Src buf
...@@ -1479,7 +1679,7 @@ struct ThreadwiseTensorSliceTransfer_StaticToStatic_InterRow ...@@ -1479,7 +1679,7 @@ struct ThreadwiseTensorSliceTransfer_StaticToStatic_InterRow
ElementwiseOperation element_op_{}; ElementwiseOperation element_op_{};
}; };
// Specilized for WMMA-Navi4 // Specialized for gfx12
template <typename SrcData, template <typename SrcData,
typename DstData, typename DstData,
typename SrcDesc, typename SrcDesc,
......
include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp
View file @ 67ab3896
...@@ -307,7 +307,7 @@ struct wmma_type<WmmaInstr::wmma_f32_16x16x16_f16_gfx12, ...@@ -307,7 +307,7 @@ struct wmma_type<WmmaInstr::wmma_f32_16x16x16_f16_gfx12,
// Wave mode dependent propety // Wave mode dependent propety
static constexpr index_t wave_size = Number<WaveSize>{}; static constexpr index_t wave_size = Number<WaveSize>{};
// * Fixed in Navi3x, Will be wave mode dependent on Navi4x // * Fixed for gfx11, Will be wave mode dependent on gfx12
// static constexpr index_t num_src_a_vgprs_per_wave = k_per_wmma / 2 * src_a_data_size / 4; // static constexpr index_t num_src_a_vgprs_per_wave = k_per_wmma / 2 * src_a_data_size / 4;
// static constexpr index_t num_src_b_vgprs_per_wave = k_per_wmma / 2 * src_b_data_size / 4; // static constexpr index_t num_src_b_vgprs_per_wave = k_per_wmma / 2 * src_b_data_size / 4;
// * num_acc_vgprs_per_wave alone M direction // * num_acc_vgprs_per_wave alone M direction
......
include/ck/utility/amd_inline_asm.hpp
View file @ 67ab3896
...@@ -4,8 +4,8 @@ ...@@ -4,8 +4,8 @@
#ifndef CK_AMD_INLINE_ASM_HPP #ifndef CK_AMD_INLINE_ASM_HPP
#define CK_AMD_INLINE_ASM_HPP #define CK_AMD_INLINE_ASM_HPP
#include "data_type.hpp"
#include "c_style_pointer_cast.hpp" #include "c_style_pointer_cast.hpp"
#include "data_type.hpp"
// TODO: deprecate all amd_assembly_outer_product_xxx // TODO: deprecate all amd_assembly_outer_product_xxx
...@@ -21,14 +21,14 @@ inline __device__ int amd_assembly_and_or_b32(int a, int b, int d) ...@@ -21,14 +21,14 @@ inline __device__ int amd_assembly_and_or_b32(int a, int b, int d)
inline __device__ half2_t amd_assembly_pk_fma_f16(half2_t a, half2_t b, half2_t c) inline __device__ half2_t amd_assembly_pk_fma_f16(half2_t a, half2_t b, half2_t c)
{ {
half2_t d; half2_t d;
asm volatile("v_pk_fma_f16 %0, %1, %2, %3;\n" : "=v"(d) : "v"(a), "v"(b), "v"(c)); asm volatile("v_pk_fma_f16 %0, %1, %2, %3" : "=v"(d) : "v"(a), "v"(b), "v"(c));
return d; return d;
} }
inline __device__ half2_t amd_assembly_pk_add_f16(half2_t a, half2_t b) inline __device__ half2_t amd_assembly_pk_add_f16(half2_t a, half2_t b)
{ {
half2_t c; half2_t c;
asm volatile("v_pk_add_f16 %0, %1, %2;\n" : "=v"(c) : "v"(a), "v"(b)); asm volatile("v_pk_add_f16 %0, %1, %2" : "=v"(c) : "v"(a), "v"(b));
return c; return c;
} }
......
include/ck/utility/data_type.hpp
View file @ 67ab3896
...@@ -19,6 +19,8 @@ struct pk_i4_t ...@@ -19,6 +19,8 @@ struct pk_i4_t
type data; type data;
__host__ __device__ constexpr pk_i4_t() : data{type{}} {} __host__ __device__ constexpr pk_i4_t() : data{type{}} {}
__host__ __device__ constexpr pk_i4_t(type init) : data{init} {} __host__ __device__ constexpr pk_i4_t(type init) : data{init} {}
__host__ __device__ constexpr operator float() const { return static_cast<int8_t>(data); }
}; };
inline constexpr auto next_pow2(uint32_t x) inline constexpr auto next_pow2(uint32_t x)
......
include/ck/utility/type_convert.hpp
View file @ 67ab3896
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -465,6 +465,19 @@ inline __host__ __device__ float2_t type_convert<float2_t, f8x2_ocp_t>(f8x2_ocp_ ...@@ -465,6 +465,19 @@ inline __host__ __device__ float2_t type_convert<float2_t, f8x2_ocp_t>(f8x2_ocp_
#endif #endif
} }
template <>
inline __host__ __device__ float2_t type_convert<float2_t, pk_i4_t>(pk_i4_t x)
{
uint8_t x_u8 = ck::bit_cast<uint8_t>(x);
uint8_t x_l = (x_u8 & 0x0f) >> 0;
uint8_t x_h = (x_u8 & 0xf0) >> 4;
auto l_f32 = ck::type_convert<float>(x_l);
auto h_f32 = ck::type_convert<float>(x_h);
return {l_f32, h_f32};
}
template <> template <>
inline __host__ __device__ half2_t type_convert<half2_t, float2_t>(float2_t x) inline __host__ __device__ half2_t type_convert<half2_t, float2_t>(float2_t x)
{ {
......
include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_kernel.hpp
View file @ 67ab3896
...@@ -47,10 +47,16 @@ struct FmhaFwdSplitKVKernel ...@@ -47,10 +47,16 @@ struct FmhaFwdSplitKVKernel
static constexpr bool kStoreLSE = FmhaPipeline::kStoreLSE; static constexpr bool kStoreLSE = FmhaPipeline::kStoreLSE;
static constexpr bool kDoFp8StaticQuant = FmhaPipeline::Problem::kDoFp8StaticQuant; static constexpr bool kDoFp8StaticQuant = FmhaPipeline::Problem::kDoFp8StaticQuant;
static constexpr bool kIsPagedKV = FmhaPipeline::Problem::kIsPagedKV; static constexpr bool kIsPagedKV = FmhaPipeline::Problem::kIsPagedKV;
static constexpr bool kMergeNumHeadGroupsSeqLenQ =
FmhaPipeline::Problem::kMergeNumHeadGroupsSeqLenQ;
using FmhaMask = ck_tile::remove_cvref_t<typename FmhaPipeline::FmhaMask>; using FmhaMask = ck_tile::remove_cvref_t<typename FmhaPipeline::FmhaMask>;
static constexpr bool kHasMask = FmhaMask::IsMasking; static constexpr bool kHasMask = FmhaMask::IsMasking;
static_assert(!kMergeNumHeadGroupsSeqLenQ ||
(kMergeNumHeadGroupsSeqLenQ && BiasEnum == BlockAttentionBiasEnum::NO_BIAS &&
!kHasMask));
// clang-format off // clang-format off
template <typename T> struct t2s; template <typename T> struct t2s;
template <> struct t2s<float> { static constexpr const char * name = "fp32"; }; template <> struct t2s<float> { static constexpr const char * name = "fp32"; };
...@@ -476,15 +482,20 @@ struct FmhaFwdSplitKVKernel ...@@ -476,15 +482,20 @@ struct FmhaFwdSplitKVKernel
} }
CK_TILE_HOST static constexpr auto GridSize(ck_tile::index_t batch_size, CK_TILE_HOST static constexpr auto GridSize(ck_tile::index_t batch_size,
ck_tile::index_t nhead, ck_tile::index_t nhead_q,
ck_tile::index_t nhead_kv,
ck_tile::index_t max_seqlen_q, ck_tile::index_t max_seqlen_q,
ck_tile::index_t hdim_v, ck_tile::index_t hdim_v,
ck_tile::index_t num_splits) ck_tile::index_t num_splits)
{ {
ck_tile::index_t nhead_ = kMergeNumHeadGroupsSeqLenQ ? nhead_kv : nhead_q;
ck_tile::index_t max_seqlen_q_ =
max_seqlen_q * (kMergeNumHeadGroupsSeqLenQ ? nhead_q / nhead_kv : 1);
// TODO: this may need tuning // TODO: this may need tuning
return dim3(ck_tile::integer_divide_ceil(max_seqlen_q, FmhaPipeline::kM0) * return dim3(ck_tile::integer_divide_ceil(max_seqlen_q_, FmhaPipeline::kM0) *
ck_tile::integer_divide_ceil(hdim_v, FmhaPipeline::kN1) * num_splits, ck_tile::integer_divide_ceil(hdim_v, FmhaPipeline::kN1) * num_splits,
nhead, nhead_,
batch_size); batch_size);
} }
...@@ -562,7 +573,7 @@ struct FmhaFwdSplitKVKernel ...@@ -562,7 +573,7 @@ struct FmhaFwdSplitKVKernel
// # of required blocks is different in each groups, terminate unnecessary blocks // # of required blocks is different in each groups, terminate unnecessary blocks
// earlier // earlier
if(kargs.seqlen_q <= i_m0) if(kargs.seqlen_q * (kMergeNumHeadGroupsSeqLenQ ? kargs.nhead_ratio_qk : 1) <= i_m0)
{ {
return; return;
} }
...@@ -617,30 +628,60 @@ struct FmhaFwdSplitKVKernel ...@@ -617,30 +628,60 @@ struct FmhaFwdSplitKVKernel
} }
// for simplicity, batch stride we just modify the pointer // for simplicity, batch stride we just modify the pointer
const index_t i_nhead_k =
(kMergeNumHeadGroupsSeqLenQ ? i_nhead : i_nhead / kargs.nhead_ratio_qk);
const QDataType* q_ptr = reinterpret_cast<const QDataType*>(kargs.q_ptr) + const QDataType* q_ptr = reinterpret_cast<const QDataType*>(kargs.q_ptr) +
static_cast<long_index_t>(i_nhead) * kargs.nhead_stride_q + static_cast<long_index_t>(i_nhead) *
(kMergeNumHeadGroupsSeqLenQ ? kargs.nhead_ratio_qk : 1) *
kargs.nhead_stride_q +
batch_offset_q; batch_offset_q;
const KDataType* k_ptr = const KDataType* k_ptr = reinterpret_cast<const KDataType*>(kargs.k_ptr) +
reinterpret_cast<const KDataType*>(kargs.k_ptr) + static_cast<long_index_t>(i_nhead_k) * kargs.nhead_stride_k +
static_cast<long_index_t>(i_nhead / kargs.nhead_ratio_qk) * kargs.nhead_stride_k + batch_offset_k;
batch_offset_k; const VDataType* v_ptr = reinterpret_cast<const VDataType*>(kargs.v_ptr) +
const VDataType* v_ptr = static_cast<long_index_t>(i_nhead_k) * kargs.nhead_stride_v +
reinterpret_cast<const VDataType*>(kargs.v_ptr) + batch_offset_v;
static_cast<long_index_t>(i_nhead / kargs.nhead_ratio_qk) * kargs.nhead_stride_v +
batch_offset_v;
ODataType* o_acc_ptr = reinterpret_cast<ODataType*>(kargs.o_acc_ptr) + ODataType* o_acc_ptr = reinterpret_cast<ODataType*>(kargs.o_acc_ptr) +
static_cast<long_index_t>(i_nhead) * kargs.nhead_stride_o_acc + static_cast<long_index_t>(i_nhead) *
(kMergeNumHeadGroupsSeqLenQ ? kargs.nhead_ratio_qk : 1) *
kargs.nhead_stride_o_acc +
batch_offset_o_acc + i_split * kargs.split_stride_o_acc; batch_offset_o_acc + i_split * kargs.split_stride_o_acc;
// Q/K/V DRAM and DRAM window // Q/K/V DRAM and DRAM window
const auto q_dram = [&]() { const auto q_dram = [&] {
const auto q_dram_naive = make_naive_tensor_view<address_space_enum::global>( const auto q_dram_naive = [&] {
q_ptr, if constexpr(kMergeNumHeadGroupsSeqLenQ)
make_tuple(kargs.seqlen_q, kargs.hdim_q), {
make_tuple(kargs.stride_q, 1), // reshape: (nhead_ratio_qk, seqlen_q, hdim_q) -> (nhead_ratio_qk * seqlen_q,
number<FmhaPipeline::kAlignmentQ>{}, // hdim_q)
number<1>{}); const auto view = make_naive_tensor_view<address_space_enum::global>(
q_ptr,
make_tuple(kargs.nhead_ratio_qk, kargs.seqlen_q, kargs.hdim_q),
make_tuple(kargs.nhead_stride_q, kargs.stride_q, 1),
number<FmhaPipeline::kAlignmentQ>{},
number<1>{});
return transform_tensor_view(
view,
make_tuple(
make_merge_transform(make_tuple(kargs.nhead_ratio_qk, kargs.seqlen_q)),
make_pass_through_transform(kargs.hdim_q)),
make_tuple(sequence<0, 1>{}, sequence<2>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
}
else
{
return make_naive_tensor_view<address_space_enum::global>(
q_ptr,
make_tuple(kargs.seqlen_q, kargs.hdim_q),
make_tuple(kargs.stride_q, 1),
number<FmhaPipeline::kAlignmentQ>{},
number<1>{});
}
}();
if constexpr(FmhaPipeline::kQLoadOnce) if constexpr(FmhaPipeline::kQLoadOnce)
{ {
return pad_tensor_view( return pad_tensor_view(
...@@ -729,7 +770,7 @@ struct FmhaFwdSplitKVKernel ...@@ -729,7 +770,7 @@ struct FmhaFwdSplitKVKernel
} }
}(); }();
auto k_page_block_navigator = [&, i_batch_ = i_batch, i_nhead_ = i_nhead]() { auto k_page_block_navigator = [&, i_batch_ = i_batch]() {
if constexpr(kIsPagedKV) if constexpr(kIsPagedKV)
{ {
const auto* block_indices = const auto* block_indices =
...@@ -739,8 +780,7 @@ struct FmhaFwdSplitKVKernel ...@@ -739,8 +780,7 @@ struct FmhaFwdSplitKVKernel
integer_divide_ceil(kv_l2p_offset + kargs.seqlen_k, kargs.page_block_size); integer_divide_ceil(kv_l2p_offset + kargs.seqlen_k, kargs.page_block_size);
const long_index_t fixed_offset = const long_index_t fixed_offset =
static_cast<long_index_t>(i_nhead_ / kargs.nhead_ratio_qk) * static_cast<long_index_t>(i_nhead_k) * kargs.nhead_stride_k;
kargs.nhead_stride_k;
return make_page_block_navigator<const KDataType, 0>( return make_page_block_navigator<const KDataType, 0>(
kargs.k_ptr, kargs.k_ptr,
...@@ -760,7 +800,7 @@ struct FmhaFwdSplitKVKernel ...@@ -760,7 +800,7 @@ struct FmhaFwdSplitKVKernel
} }
}(); }();
auto v_page_block_navigator = [&, i_batch_ = i_batch, i_nhead_ = i_nhead]() { auto v_page_block_navigator = [&, i_batch_ = i_batch]() {
if constexpr(kIsPagedKV) if constexpr(kIsPagedKV)
{ {
const auto* block_indices = const auto* block_indices =
...@@ -770,8 +810,7 @@ struct FmhaFwdSplitKVKernel ...@@ -770,8 +810,7 @@ struct FmhaFwdSplitKVKernel
integer_divide_ceil(kv_l2p_offset + kargs.seqlen_k, kargs.page_block_size); integer_divide_ceil(kv_l2p_offset + kargs.seqlen_k, kargs.page_block_size);
const long_index_t fixed_offset = const long_index_t fixed_offset =
static_cast<long_index_t>(i_nhead_ / kargs.nhead_ratio_qk) * static_cast<long_index_t>(i_nhead_k) * kargs.nhead_stride_v;
kargs.nhead_stride_v;
return make_page_block_navigator<const VDataType, 1>( return make_page_block_navigator<const VDataType, 1>(
kargs.v_ptr, kargs.v_ptr,
...@@ -842,19 +881,40 @@ struct FmhaFwdSplitKVKernel ...@@ -842,19 +881,40 @@ struct FmhaFwdSplitKVKernel
// lse acc // lse acc
auto lse_acc_dram_window = [&, i_nhead_ = i_nhead, i_split_ = i_split]() { auto lse_acc_dram_window = [&, i_nhead_ = i_nhead, i_split_ = i_split]() {
constexpr auto lse_acc_dram_window_lengths = make_tuple(number<FmhaPipeline::kM0>{}); constexpr auto lse_acc_dram_window_lengths = make_tuple(number<FmhaPipeline::kM0>{});
LSEDataType* lse_acc_ptr = LSEDataType* lse_acc_ptr = reinterpret_cast<LSEDataType*>(kargs.lse_acc_ptr) +
reinterpret_cast<LSEDataType*>(kargs.lse_acc_ptr) + static_cast<long_index_t>(i_nhead_) *
static_cast<long_index_t>(i_nhead_) * kargs.nhead_stride_lse_acc + (kMergeNumHeadGroupsSeqLenQ ? kargs.nhead_ratio_qk : 1) *
batch_offset_lse_acc + i_split_ * kargs.split_stride_lse_acc; kargs.nhead_stride_lse_acc +
batch_offset_lse_acc + i_split_ * kargs.split_stride_lse_acc;
const auto lse_acc_dram = [&]() {
const auto lse_acc_dram_naive = const auto lse_acc_dram = [&] {
make_naive_tensor_view<address_space_enum::global>(lse_acc_ptr, const auto lse_acc_dram_naive = [&] {
make_tuple(kargs.seqlen_q), if constexpr(kMergeNumHeadGroupsSeqLenQ)
make_tuple(1), {
number<1>{}, // reshape: (nhead_ratio_qk, seqlen_q) -> (nhead_ratio_qk * seqlen_q)
number<1>{}); const auto view = make_naive_tensor_view<address_space_enum::global>(
lse_acc_ptr,
make_tuple(kargs.nhead_ratio_qk, kargs.seqlen_q),
make_tuple(kargs.nhead_stride_lse_acc, 1),
number<1>{},
number<1>{});
return transform_tensor_view(view,
make_tuple(make_merge_transform(make_tuple(
kargs.nhead_ratio_qk, kargs.seqlen_q))),
make_tuple(sequence<0, 1>{}),
make_tuple(sequence<0>{}));
}
else
{
return make_naive_tensor_view<address_space_enum::global>(
lse_acc_ptr,
make_tuple(kargs.seqlen_q),
make_tuple(1),
number<1>{},
number<1>{});
}
}();
return pad_tensor_view( return pad_tensor_view(
lse_acc_dram_naive, lse_acc_dram_window_lengths, sequence<kPadSeqLenQ>{}); lse_acc_dram_naive, lse_acc_dram_window_lengths, sequence<kPadSeqLenQ>{});
}(); }();
...@@ -953,13 +1013,37 @@ struct FmhaFwdSplitKVKernel ...@@ -953,13 +1013,37 @@ struct FmhaFwdSplitKVKernel
}(); }();
// Oacc DRAM and Oacc DRAM window // Oacc DRAM and Oacc DRAM window
auto o_acc_dram = [&]() { auto o_acc_dram = [&] {
const auto o_acc_dram_naive = make_naive_tensor_view<address_space_enum::global>( const auto o_acc_dram_naive = [&] {
o_acc_ptr, if constexpr(kMergeNumHeadGroupsSeqLenQ)
make_tuple(kargs.seqlen_q, kargs.hdim_v), {
make_tuple(kargs.stride_o_acc, 1), // reshape: (nhead_ratio_qk, seqlen_q, hdim_v) -> (nhead_ratio_qk * seqlen_q,
number<FmhaPipeline::kAlignmentOacc>{}, // hdim_v)
number<1>{}); const auto view = make_naive_tensor_view<address_space_enum::global>(
o_acc_ptr,
make_tuple(kargs.nhead_ratio_qk, kargs.seqlen_q, kargs.hdim_v),
make_tuple(kargs.nhead_stride_o_acc, kargs.stride_o_acc, 1),
number<FmhaPipeline::kAlignmentOacc>{},
number<1>{});
return transform_tensor_view(
view,
make_tuple(
make_merge_transform(make_tuple(kargs.nhead_ratio_qk, kargs.seqlen_q)),
make_pass_through_transform(kargs.hdim_v)),
make_tuple(sequence<0, 1>{}, sequence<2>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
}
else
{
return make_naive_tensor_view<address_space_enum::global>(
o_acc_ptr,
make_tuple(kargs.seqlen_q, kargs.hdim_v),
make_tuple(kargs.stride_o_acc, 1),
number<FmhaPipeline::kAlignmentOacc>{},
number<1>{});
}
}();
return pad_tensor_view( return pad_tensor_view(
o_acc_dram_naive, o_acc_dram_naive,
......
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_problem.hpp
View file @ 67ab3896
...@@ -94,16 +94,17 @@ struct BlockFmhaFwdSplitKVPipelineProblem ...@@ -94,16 +94,17 @@ struct BlockFmhaFwdSplitKVPipelineProblem
static constexpr bool kIsGroupMode = kIsGroupMode_; static constexpr bool kIsGroupMode = kIsGroupMode_;
// attributes from traits // attributes from traits
static constexpr bool kPadSeqLenQ = Traits::kPadSeqLenQ; static constexpr bool kPadSeqLenQ = Traits::kPadSeqLenQ;
static constexpr bool kPadSeqLenK = Traits::kPadSeqLenK; static constexpr bool kPadSeqLenK = Traits::kPadSeqLenK;
static constexpr bool kPadHeadDimQ = Traits::kPadHeadDimQ; static constexpr bool kPadHeadDimQ = Traits::kPadHeadDimQ;
static constexpr bool kPadHeadDimV = Traits::kPadHeadDimV; static constexpr bool kPadHeadDimV = Traits::kPadHeadDimV;
static constexpr auto BiasEnum = Traits::BiasEnum; static constexpr auto BiasEnum = Traits::BiasEnum;
static constexpr bool kStoreLSE = Traits::kStoreLSE; static constexpr bool kStoreLSE = Traits::kStoreLSE;
static constexpr bool kDoFp8StaticQuant = Traits::kDoFp8StaticQuant; static constexpr bool kDoFp8StaticQuant = Traits::kDoFp8StaticQuant;
static constexpr bool kIsPagedKV = Traits::kIsPagedKV; static constexpr bool kIsPagedKV = Traits::kIsPagedKV;
static constexpr bool kHasUnevenSplits = kIsGroupMode || Traits::kHasUnevenSplits; static constexpr bool kHasUnevenSplits = kIsGroupMode || Traits::kHasUnevenSplits;
static constexpr index_t kBlockPerCu = Traits::kBlockPerCu; static constexpr bool kMergeNumHeadGroupsSeqLenQ = Traits::kMergeNumHeadGroupsSeqLenQ;
static constexpr index_t kBlockPerCu = Traits::kBlockPerCu;
}; };
// extract tile size attributes to remove dependency on traits // extract tile size attributes to remove dependency on traits
......
include/ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp
View file @ 67ab3896
...@@ -43,7 +43,8 @@ template <bool kPadSeqLenQ_ /* padding for seqlen_q */, ...@@ -43,7 +43,8 @@ template <bool kPadSeqLenQ_ /* padding for seqlen_q */,
bool kDoFp8StaticQuant_, bool kDoFp8StaticQuant_,
bool kIsPagedKV_, bool kIsPagedKV_,
bool kHasUnevenSplits_, bool kHasUnevenSplits_,
index_t kBlockPerCu_ = -1 /* overwrite occupancy if not -1 */> bool kMergeNumHeadGroupsSeqLenQ_ = false,
index_t kBlockPerCu_ = -1 /* overwrite occupancy if not -1 */>
struct TileFmhaFwdSplitKVTraits struct TileFmhaFwdSplitKVTraits
{ {
static constexpr bool kPadSeqLenQ = kPadSeqLenQ_; static constexpr bool kPadSeqLenQ = kPadSeqLenQ_;
...@@ -56,8 +57,9 @@ struct TileFmhaFwdSplitKVTraits ...@@ -56,8 +57,9 @@ struct TileFmhaFwdSplitKVTraits
static constexpr bool kDoFp8StaticQuant = kDoFp8StaticQuant_; static constexpr bool kDoFp8StaticQuant = kDoFp8StaticQuant_;
static constexpr bool kIsPagedKV = kIsPagedKV_; static constexpr bool kIsPagedKV = kIsPagedKV_;
// determine if some split (length) is not divisible by tile size // determine if some split (length) is not divisible by tile size
static constexpr bool kHasUnevenSplits = kHasUnevenSplits_; static constexpr bool kHasUnevenSplits = kHasUnevenSplits_;
static constexpr index_t kBlockPerCu = kBlockPerCu_; static constexpr bool kMergeNumHeadGroupsSeqLenQ = kMergeNumHeadGroupsSeqLenQ_;
static constexpr index_t kBlockPerCu = kBlockPerCu_;
}; };
template <bool kPadSeqLenQ_ /* padding for seqlen_q */, template <bool kPadSeqLenQ_ /* padding for seqlen_q */,
......
include/ck_tile/ops/layernorm2d/kernel/layernorm2d_fwd_kernel.hpp
View file @ 67ab3896
...@@ -15,6 +15,7 @@ struct Layernorm2dFwdHostArgs ...@@ -15,6 +15,7 @@ struct Layernorm2dFwdHostArgs
const void* p_x; // [m ,n], input, fp16/bf16 const void* p_x; // [m ,n], input, fp16/bf16
const void* p_x_residual; // [m ,n], shortcut input, prec same as input, nullptr if not used const void* p_x_residual; // [m ,n], shortcut input, prec same as input, nullptr if not used
const void* p_x_scale; // [1 ,n], smooth scale input, fp32, nullptr if not used const void* p_x_scale; // [1 ,n], smooth scale input, fp32, nullptr if not used
const void* p_x_bias; // [1, n], bias, prec same as input
const void* p_gamma; // [1, n], gamma, prec same as input const void* p_gamma; // [1, n], gamma, prec same as input
const void* p_beta; // [1, n], beta, prec same as input const void* p_beta; // [1, n], beta, prec same as input
...@@ -43,6 +44,7 @@ struct Layernorm2dFwd ...@@ -43,6 +44,7 @@ struct Layernorm2dFwd
using Problem = typename Pipeline::Problem; using Problem = typename Pipeline::Problem;
using XDataType = remove_cvref_t<typename Problem::XDataType>; using XDataType = remove_cvref_t<typename Problem::XDataType>;
using XBiasDataType = remove_cvref_t<typename Problem::XBiasDataType>;
using GammaDataType = remove_cvref_t<typename Problem::GammaDataType>; using GammaDataType = remove_cvref_t<typename Problem::GammaDataType>;
using BetaDataType = remove_cvref_t<typename Problem::BetaDataType>; using BetaDataType = remove_cvref_t<typename Problem::BetaDataType>;
using ComputeDataType = remove_cvref_t<typename Problem::ComputeDataType>; using ComputeDataType = remove_cvref_t<typename Problem::ComputeDataType>;
...@@ -67,6 +69,7 @@ struct Layernorm2dFwd ...@@ -67,6 +69,7 @@ struct Layernorm2dFwd
static constexpr bool kPadM = false; // always no need to pad along M static constexpr bool kPadM = false; // always no need to pad along M
static constexpr bool kPadN = Problem::Traits::kPadN; static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr bool kTwoPass = Problem::Traits::kTwoPass; static constexpr bool kTwoPass = Problem::Traits::kTwoPass;
static constexpr auto kXbias = Problem::Traits::kXbias;
static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd; static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd;
static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant; static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant;
...@@ -82,6 +85,7 @@ struct Layernorm2dFwd ...@@ -82,6 +85,7 @@ struct Layernorm2dFwd
const void* p_x; // [m ,n], input, fp16/bf16 const void* p_x; // [m ,n], input, fp16/bf16
const void* p_x_residual; // [m ,n], shortcut input, prec same as input, nullptr if not used const void* p_x_residual; // [m ,n], shortcut input, prec same as input, nullptr if not used
const void* p_x_scale; // [1 ,n], smooth scale input, fp32, nullptr if not used const void* p_x_scale; // [1 ,n], smooth scale input, fp32, nullptr if not used
const void* p_x_bias; // [1, n], bias, prec same as input
const void* p_gamma; // [1, n], gamma, prec same as input const void* p_gamma; // [1, n], gamma, prec same as input
const void* p_beta; // [1, n], beta, prec same as input const void* p_beta; // [1, n], beta, prec same as input
...@@ -108,6 +112,7 @@ struct Layernorm2dFwd ...@@ -108,6 +112,7 @@ struct Layernorm2dFwd
return Kargs{hargs.p_x, return Kargs{hargs.p_x,
hargs.p_x_residual, hargs.p_x_residual,
hargs.p_x_scale, hargs.p_x_scale,
hargs.p_x_bias,
hargs.p_gamma, hargs.p_gamma,
hargs.p_beta, hargs.p_beta,
hargs.p_y, hargs.p_y,
...@@ -152,6 +157,7 @@ struct Layernorm2dFwd ...@@ -152,6 +157,7 @@ struct Layernorm2dFwd
using S_ = typename Problem::BlockShape; using S_ = typename Problem::BlockShape;
auto surfix = [&] () { auto surfix = [&] () {
std::string n; std::string n;
if (kXbias != Layernorm2dXBiasEnum::NO_BIAS) n += _SS_("_") + Layernorm2dXBiasEnumName<kXbias>::name;
if (kFusedAdd != Layernorm2dFusedAddEnum::NO_ADD) n += _SS_("_") + Layernorm2dFusedAddEnumName<kFusedAdd>::name; if (kFusedAdd != Layernorm2dFusedAddEnum::NO_ADD) n += _SS_("_") + Layernorm2dFusedAddEnumName<kFusedAdd>::name;
if (kFusedQuant != Layernorm2dFusedQuantEnum::NO_SWEEP) n += _SS_("_") + Layernorm2dFusedQuantEnumName<kFusedQuant>::name; if (kFusedQuant != Layernorm2dFusedQuantEnum::NO_SWEEP) n += _SS_("_") + Layernorm2dFusedQuantEnumName<kFusedQuant>::name;
if (kPadN) n += "_pn"; if (kPadN) n += "_pn";
...@@ -228,6 +234,27 @@ struct Layernorm2dFwd ...@@ -228,6 +234,27 @@ struct Layernorm2dFwd
} }
}(); }();
const auto x_bias_window = [&]() {
if constexpr(kXbias == Layernorm2dXBiasEnum::ADD_BIAS)
{
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<const XBiasDataType*>(kargs.p_x_bias),
make_tuple(kargs.n),
make_tuple(1),
number<Vector_N>{},
number<1>{});
const auto tmp2_ =
pad_tensor_view(tmp_, make_tuple(number<Block_N>{}), sequence<false>{});
return make_tile_window(tmp2_, make_tuple(number<Block_N>{}), {0});
}
else
{
return make_null_tile_window(make_tuple(number<Block_N>{}));
}
}();
const auto gamma_window = [&]() { const auto gamma_window = [&]() {
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>( const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<const GammaDataType*>(kargs.p_gamma), static_cast<const GammaDataType*>(kargs.p_gamma),
...@@ -371,6 +398,7 @@ struct Layernorm2dFwd ...@@ -371,6 +398,7 @@ struct Layernorm2dFwd
Pipeline{}(x_window, Pipeline{}(x_window,
x_residual_window, x_residual_window,
x_bias_window,
gamma_window, gamma_window,
beta_window, beta_window,
y_window, y_window,
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_default_policy.hpp
View file @ 67ab3896
...@@ -4,8 +4,8 @@ ...@@ -4,8 +4,8 @@
#pragma once #pragma once
#include "ck_tile/core.hpp" #include "ck_tile/core.hpp"
#include "ck_tile/ops/welford/block/block_welford_problem.hpp" #include "ck_tile/ops/norm_reduce/block/block_norm_reduce_problem.hpp"
#include "ck_tile/ops/welford/block/block_welford.hpp" #include "ck_tile/ops/norm_reduce/block/block_norm_reduce.hpp"
namespace ck_tile { namespace ck_tile {
...@@ -43,36 +43,38 @@ struct Layernorm2dFwdPipelineDefaultPolicy ...@@ -43,36 +43,38 @@ struct Layernorm2dFwdPipelineDefaultPolicy
} }
template <typename Problem> template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockWelford() CK_TILE_HOST_DEVICE static constexpr auto GetBlockNormReduce()
{ {
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType, using P_ = BlockNormReduceProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType, typename Problem::ComputeDataType,
typename Problem::BlockShape, typename Problem::BlockShape,
Problem::Traits::kFastFDiv>; Problem::Traits::kFastFDiv,
Problem::Traits::kWelford>;
return BlockWelford<P_>{}; return BlockNormReduce<P_>{};
} }
template <typename Problem> template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockWelfordSync() CK_TILE_HOST_DEVICE static constexpr auto GetBlockNormReduceSync()
{ {
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType, using P_ = BlockNormReduceProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType, typename Problem::ComputeDataType,
typename Problem::BlockShape, typename Problem::BlockShape,
Problem::Traits::kFastFDiv>; Problem::Traits::kFastFDiv,
Problem::Traits::kWelford>;
return BlockWelfordSync<P_>{}; return BlockNormReduceSync<P_>{};
} }
template <typename Problem> template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockWelfordCrossWarpSync() CK_TILE_HOST_DEVICE static constexpr auto GetBlockNormReduceCrossWarpSync()
{ {
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType, using P_ = BlockNormReduceProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType, typename Problem::ComputeDataType,
typename Problem::BlockShape, typename Problem::BlockShape,
Problem::Traits::kFastFDiv>; Problem::Traits::kFastFDiv,
Problem::Traits::kWelford>;
return BlockWelfordCrossWarpSync<P_>{}; return BlockNormReduceCrossWarpSync<P_>{};
} }
template <typename Problem> template <typename Problem>
...@@ -80,19 +82,20 @@ struct Layernorm2dFwdPipelineDefaultPolicy ...@@ -80,19 +82,20 @@ struct Layernorm2dFwdPipelineDefaultPolicy
{ {
if constexpr(Problem::kNeedCrossWarpSync) if constexpr(Problem::kNeedCrossWarpSync)
{ {
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType, using P_ = BlockNormReduceProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType, typename Problem::ComputeDataType,
typename Problem::BlockShape, typename Problem::BlockShape,
Problem::Traits::kFastFDiv>; Problem::Traits::kFastFDiv,
Problem::Traits::kWelford>;
using block_welford = BlockWelford<P_>; using block_welford = BlockNormReduce<P_>;
using x_block_tile = using x_block_tile =
decltype(make_static_distributed_tensor<typename Problem::ComputeDataType>( decltype(make_static_distributed_tensor<typename Problem::ComputeDataType>(
MakeXBlockTileDistribution<Problem>())); MakeXBlockTileDistribution<Problem>()));
using mean_var_block_tile = using mean_var_block_tile =
decltype(block_welford::template MakeMeanVarBlockTile<x_block_tile>()); decltype(block_welford::template MakeMeanVarBlockTile<x_block_tile>());
return GetBlockWelfordCrossWarpSync<Problem>() return GetBlockNormReduceCrossWarpSync<Problem>()
.template GetSmemSize<mean_var_block_tile>(); .template GetSmemSize<mean_var_block_tile>();
} }
else else
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_one_pass.hpp
View file @ 67ab3896
...@@ -18,6 +18,7 @@ struct Layernorm2dFwdPipelineOnePass ...@@ -18,6 +18,7 @@ struct Layernorm2dFwdPipelineOnePass
using Policy = ck_tile::remove_cvref_t<Policy_>; using Policy = ck_tile::remove_cvref_t<Policy_>;
using XDataType = ck_tile::remove_cvref_t<typename Problem::XDataType>; using XDataType = ck_tile::remove_cvref_t<typename Problem::XDataType>;
using XBiasDataType = ck_tile::remove_cvref_t<typename Problem::XBiasDataType>;
using GammaDataType = ck_tile::remove_cvref_t<typename Problem::GammaDataType>; using GammaDataType = ck_tile::remove_cvref_t<typename Problem::GammaDataType>;
using BetaDataType = ck_tile::remove_cvref_t<typename Problem::BetaDataType>; using BetaDataType = ck_tile::remove_cvref_t<typename Problem::BetaDataType>;
using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>; using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
...@@ -37,6 +38,8 @@ struct Layernorm2dFwdPipelineOnePass ...@@ -37,6 +38,8 @@ struct Layernorm2dFwdPipelineOnePass
static constexpr bool kPadM = false; // TODO - BlockLayernorm2dFwdProblem::kPadM static constexpr bool kPadM = false; // TODO - BlockLayernorm2dFwdProblem::kPadM
static constexpr bool kPadN = Problem::Traits::kPadN; static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr bool kFastFDiv = Problem::Traits::kFastFDiv; static constexpr bool kFastFDiv = Problem::Traits::kFastFDiv;
static constexpr bool kWelford = Problem::Traits::kWelford;
static constexpr auto kXbias = Problem::Traits::kXbias;
static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd; static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd;
static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant; static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant;
...@@ -54,6 +57,7 @@ struct Layernorm2dFwdPipelineOnePass ...@@ -54,6 +57,7 @@ struct Layernorm2dFwdPipelineOnePass
template <typename XWindow, template <typename XWindow,
typename XResidualWindow, typename XResidualWindow,
typename XBiasWindow,
typename GammaWindow, typename GammaWindow,
typename BetaWindow, typename BetaWindow,
typename YWindow, typename YWindow,
...@@ -65,6 +69,7 @@ struct Layernorm2dFwdPipelineOnePass ...@@ -65,6 +69,7 @@ struct Layernorm2dFwdPipelineOnePass
typename Epilogue> typename Epilogue>
CK_TILE_DEVICE auto operator()(const XWindow& x_window_, CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
const XResidualWindow& x_residual_window_, const XResidualWindow& x_residual_window_,
const XBiasWindow& x_bias_window_,
const GammaWindow& gamma_window_, const GammaWindow& gamma_window_,
const BetaWindow& beta_window_, const BetaWindow& beta_window_,
YWindow& y_window_, YWindow& y_window_,
...@@ -80,6 +85,8 @@ struct Layernorm2dFwdPipelineOnePass ...@@ -80,6 +85,8 @@ struct Layernorm2dFwdPipelineOnePass
{ {
const auto x_window = const auto x_window =
make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>()); make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
const auto x_bias_window = make_tile_window(
x_bias_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>());
const auto gamma_window = make_tile_window( const auto gamma_window = make_tile_window(
gamma_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>()); gamma_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>());
const auto beta_window = make_tile_window( const auto beta_window = make_tile_window(
...@@ -89,23 +96,38 @@ struct Layernorm2dFwdPipelineOnePass ...@@ -89,23 +96,38 @@ struct Layernorm2dFwdPipelineOnePass
auto y_residual_window = make_tile_window( auto y_residual_window = make_tile_window(
y_residual_window_, Policy::template MakeXBlockTileDistribution<Problem>()); y_residual_window_, Policy::template MakeXBlockTileDistribution<Problem>());
auto x = load_tile(x_window); auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window); auto x_resi = load_tile(x_residual_window);
const auto x_bias = load_tile(x_bias_window);
int cur_count = 0; int cur_count = 0;
int max_count = int max_count =
block_tile_welford_calculate_max_count<typename Problem::BlockShape>(row_size); block_tile_welford_calculate_max_count<typename Problem::BlockShape>(row_size);
auto block_welford = Policy::template GetBlockWelford<Problem>(); auto block_norm_reduce = Policy::template GetBlockNormReduce<Problem>();
auto block_welford_sync = Policy::template GetBlockWelfordSync<Problem>(); auto block_norm_reduce_sync = Policy::template GetBlockNormReduceSync<Problem>();
auto block_welford_cross_warp_sync = auto block_norm_reduce_cross_warp_sync =
Policy::template GetBlockWelfordCrossWarpSync<Problem>(); Policy::template GetBlockNormReduceCrossWarpSync<Problem>();
using XTensorType = decltype(cast_tile<ComputeDataType>(x));
auto mean = block_norm_reduce.template MakeMeanVarBlockTile<XTensorType>();
auto var = block_norm_reduce.template MakeMeanVarBlockTile<XTensorType>();
clear_tile(mean);
clear_tile(var);
// load gamma/beta (TODO: support no gamma/beta?) // load gamma/beta (TODO: support no gamma/beta?)
const auto gamma = load_tile(gamma_window); const auto gamma = load_tile(gamma_window);
const auto beta = load_tile(beta_window); const auto beta = load_tile(beta_window);
auto acc = cast_tile<ComputeDataType>(x); auto acc = cast_tile<ComputeDataType>(x);
if constexpr(kXbias == Layernorm2dXBiasEnum::ADD_BIAS)
{
sweep_tile(x, [&](auto idx) {
// compute x = bias + x
constexpr auto j_idx = make_tuple(idx[number<1>{}]);
acc(idx) = type_convert<ComputeDataType>(x_bias[j_idx]) + acc(idx);
});
}
if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE || if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE ||
kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD) kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD)
{ {
...@@ -117,12 +139,21 @@ struct Layernorm2dFwdPipelineOnePass ...@@ -117,12 +139,21 @@ struct Layernorm2dFwdPipelineOnePass
store_tile(y_residual_window, cast_tile<YResidualDataType>(acc)); store_tile(y_residual_window, cast_tile<YResidualDataType>(acc));
} }
// compute welford each-thread->cross-lane->cross-warp // compute reduce each-thread->cross-lane->cross-warp
auto [mean, var] = block_welford(acc, cur_count, max_count); block_norm_reduce(acc, mean, var, cur_count, max_count);
block_welford_sync(mean, var, cur_count); block_norm_reduce_sync(mean, var, cur_count);
block_welford_cross_warp_sync(mean, var, cur_count, smem); block_norm_reduce_cross_warp_sync(mean, var, cur_count, smem);
block_tile_welford_post_scale_var(var, cur_count, constant<kFastFDiv>{}); if(kWelford)
{
block_tile_welford_post_scale_var(var, cur_count, constant<kFastFDiv>{});
}
else
{
sweep_tile(mean, [&](auto idx) {
mean(idx) = mean(idx) / type_convert<MeanDataType>(row_size);
var(idx) = var(idx) / type_convert<MeanDataType>(row_size) - mean(idx) * mean(idx);
});
}
// compute inv-std // compute inv-std
auto inv_std = tile_elementwise_in( auto inv_std = tile_elementwise_in(
[&](const auto& v_) { [&](const auto& v_) {
...@@ -153,8 +184,7 @@ struct Layernorm2dFwdPipelineOnePass ...@@ -153,8 +184,7 @@ struct Layernorm2dFwdPipelineOnePass
const auto beta_ = type_convert<ComputeDataType>(beta[j_idx]); const auto beta_ = type_convert<ComputeDataType>(beta[j_idx]);
auto ln_ = (acc[idx] - mean_[i_idx]) * inv_std[i_idx] * gamma_ + beta_; auto ln_ = (acc[idx] - mean_[i_idx]) * inv_std[i_idx] * gamma_ + beta_;
ln(idx) = ln_;
ln(idx) = ln_;
}); });
if constexpr(kFusedQuant == Layernorm2dFusedQuantEnum::DYNAMIC_QUANT || if constexpr(kFusedQuant == Layernorm2dFusedQuantEnum::DYNAMIC_QUANT ||
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_problem.hpp
View file @ 67ab3896
...@@ -8,6 +8,7 @@ ...@@ -8,6 +8,7 @@
namespace ck_tile { namespace ck_tile {
template <typename XDataType_, template <typename XDataType_,
typename XBiasDataType_,
typename GammaDataType_, typename GammaDataType_,
typename BetaDataType_, typename BetaDataType_,
typename ComputeDataType_, typename ComputeDataType_,
...@@ -21,6 +22,7 @@ template <typename XDataType_, ...@@ -21,6 +22,7 @@ template <typename XDataType_,
struct Layernorm2dFwdPipelineProblem struct Layernorm2dFwdPipelineProblem
{ {
using XDataType = remove_cvref_t<XDataType_>; using XDataType = remove_cvref_t<XDataType_>;
using XBiasDataType = remove_cvref_t<XBiasDataType_>;
using GammaDataType = remove_cvref_t<GammaDataType_>; using GammaDataType = remove_cvref_t<GammaDataType_>;
using BetaDataType = remove_cvref_t<BetaDataType_>; using BetaDataType = remove_cvref_t<BetaDataType_>;
using ComputeDataType = remove_cvref_t<ComputeDataType_>; using ComputeDataType = remove_cvref_t<ComputeDataType_>;
......
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