Skip to content

GitLab

Commit bbe74503 authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'develop' into amd-develop

parents 8b76b832 f53ede26
Hide whitespace changes
Inline Side-by-side
Showing with 819 additions and 199 deletions
include/ck/tensor_operation/gpu/element/quantization_operation.hpp
View file @ bbe74503
#pragma once
#include "ck/utility/data_type.hpp"
// #include "ck/utility/get_id.hpp"
namespace ck {
namespace tensor_operation {
namespace element_wise {
// Y = Sy * Qy
// W = Sw * Qw
// X = Sx * Qx
// B = Sb * Qb = Sw * Sx * Qb
// Where X, W, Y are float32, Qx, Qw, Qy are int8
// Sx, Sw, Sy are scale of x, w, y (float32), which is calculated from quantization range
// Qb is int32, scale of B is Sw * Sx for convenient
// Y = W @ X, where @ is convolution or matrix multiplication
// Sy * Qy = Sw * Qw @ Sx * Qx
// Qy = [(Sw*Sx)/Sy] * Qw @ Qx
// For Activation function which is piecewise linear function, such as relu, leaky relu ...etc
// Activation(Sy * Qy) = Sy * Activation(Qy)
template <typename Activation>
struct Activation_Mul_Clamp
{
// Convolution + Activation (piecewise linear function)
// If an activation is piecewise linear function, then Activation(Sy * Qy) = Sy * Activation(Qy)
// Z = Activation(Y) = Activation(W @ X)
// Sz * Qz = Activation(Sy * Qy)
// Qz = Sy / Sz * Activation(Qy) = (Sw * Sx / Sz) * Activation(Qw @ Qx)
// requantScale_ = Sw * Sx / Sz
Activation_Mul_Clamp(float requantScale, Activation activationOp)
: requantScale_(requantScale), activationOp_(activationOp)
{
......@@ -17,26 +38,66 @@ struct Activation_Mul_Clamp
__host__ __device__ constexpr void operator()(int8_t& y, const int32_t& x) const
{
float x_fp32 = ck::type_convert<float>(x);
activationOp_(x_fp32, x_fp32);
float y_fp32 = math::clamp(requantScale_ * x_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
float y_fp32 = ck::type_convert<float>(x);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
__device__ constexpr void operator()(int32_t& y, const int32_t& x) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
__host__ __device__ constexpr void operator()(float& y, const int32_t& x) const
__host__ constexpr void operator()(float& y, const float& x) const
{
// We might type_convert to int8 after lambda in someplace
float x_fp32 = ck::type_convert<float>(x);
activationOp_(x_fp32, x_fp32);
y = math::clamp(requantScale_ * x_fp32, -128.f, 127.f);
// CAUSION - We might float in & float out in reference code
activationOp_(y, x);
y = math::clamp(requantScale_ * y, -128.f, 127.f);
}
float requantScale_;
Activation activationOp_;
};
// For Activation function which is non piecewise linear function, such as TanH, Sigmoid ...etc
// If an activation is not piecewise linear function
// then Activation(Sy * Qy) != Sy * Activation(Qy)
template <typename Activation>
struct Mul_Activation_Mul_Clamp
{
// Convolution + Activation (non piecewise linear function)
// Z = Activation(Y) = Activation(W @ X)
// Sz * Qz = Activation(Sy * Qy)
// Qz = S1 * Activation[Sacc * (Qw @ Qx)]
// Where S1 = 1 / Sz, Sacc = Sw * Sx
Mul_Activation_Mul_Clamp(float scale_z_inv, float scaleAcc, Activation activationOp)
: scale_z_inv_(scale_z_inv), scaleAcc_(scaleAcc), activationOp_(activationOp)
{
}
__host__ __device__ constexpr void operator()(int8_t& y, const int32_t& x) const
{
float y_fp32 = ck::type_convert<float>(x);
y_fp32 = scaleAcc_ * y_fp32;
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
float scale_z_inv_;
float scaleAcc_;
Activation activationOp_;
};
// Conv Perchannel quantization + Activation function which is piecewise linear function, such as
// relu, leaky relu ...etc
// Activation(Sy * Qy) = Sy * Activation(Qy)
template <typename Activation>
struct Activation_Mul2_Clamp
{
......@@ -51,13 +112,35 @@ struct Activation_Mul2_Clamp
y = ck::type_convert<int8_t>(y_fp32);
}
__device__ constexpr void
operator()(int32_t& y, const int32_t& x, const float& requantScale) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
Activation activationOp_;
};
// For Activation function which is piecewise linear function, such as relu, leaky relu ...etc
// Activation(Sy * Qy) = Sy * Activation(Qy)
template <typename Activation>
struct Add_Activation_Mul_Clamp
{
// Convolution + bias
// Let Bias = B = Sw * Sx * Qb
// Where Qb is int32
// Y = W @ X + B
// Sy * Qy = Sw * Qw @ Sx * Qx + Sw * Sx * Qb
// Qy = [(Sw*Sx)/Sy] * (Qw @ Qx + Qb)
// For activation, Z = Activaiton(Y)
// Sz * Qz = Activation(Sy * Qy)
// Qz = Sy / Sz * Activation(Qy) = [(Sw*Sx)/Sz] * Activation(Qw @ Qx + Qb)
Add_Activation_Mul_Clamp(float requantScale, Activation activationOp)
: requantScale_(requantScale), activationOp_(activationOp)
{
......@@ -72,6 +155,17 @@ struct Add_Activation_Mul_Clamp
y = ck::type_convert<int8_t>(y_fp32);
}
__host__ __device__ constexpr void
operator()(int32_t& y, const int32_t& x, const int32_t& bias) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x + bias);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
float requantScale_;
Activation activationOp_;
};
......@@ -92,15 +186,33 @@ struct Add_Activation_Mul2_Clamp
y = ck::type_convert<int8_t>(y_fp32);
}
__host__ __device__ constexpr void
operator()(int32_t& y, const int32_t& x, const int32_t& bias, const float& requantScale) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x + bias);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
Activation activationOp_;
};
// For Activation function which is non piecewise linear function, such as TanH, Sigmoid ...etc
// If an activation is not piecewise linear function
// then Activation(Sy * Qy) != Sy * Activation(Qy)
template <typename Activation>
struct Add_Mul_Activation_Mul_Clamp
{
Add_Mul_Activation_Mul_Clamp(float requantScale1, float requantScale2, Activation activationOp)
: requantScale1_(requantScale1), requantScale2_(requantScale2), activationOp_(activationOp)
// Convolution + Activation (non piecewise linear function)
// Z = Activation(Y) = Activation(W @ X + B)
// Sz * Qz = Activation(Sy * Qy)
// Qz = S1 * Activation[Sacc * (Qw @ Qx + Qb)]
// Where S1 = 1 / Sz, Sacc = Sw * Sx
Add_Mul_Activation_Mul_Clamp(float scale_z_inv, float scaleAcc, Activation activationOp)
: scale_z_inv_(scale_z_inv), scaleAcc_(scaleAcc), activationOp_(activationOp)
{
}
......@@ -108,14 +220,64 @@ struct Add_Mul_Activation_Mul_Clamp
operator()(int8_t& y, const int32_t& x, const int32_t& bias) const
{
float y_fp32 = ck::type_convert<float>(x + bias);
y_fp32 = requantScale1_ * y_fp32;
y_fp32 = scaleAcc_ * y_fp32;
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
__host__ __device__ constexpr void
operator()(int32_t& y, const int32_t& x, const int32_t& bias) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x + bias);
y_fp32 = scaleAcc_ * y_fp32;
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
float scale_z_inv_;
float scaleAcc_;
Activation activationOp_;
};
// Conv Perchannel quantization + Activation function which is non piecewise linear function,
// such as TanH, Sigmoid ...etc
// If an activation is not piecewise linear function
// then Activation(Sy *Qy) != Sy * Activation(Qy)
template <typename Activation>
struct Add_Mul2_Activation_Mul_Clamp
{
Add_Mul2_Activation_Mul_Clamp(float scale_z_inv, Activation activationOp)
: scale_z_inv_(scale_z_inv), activationOp_(activationOp)
{
}
__host__ __device__ constexpr void
operator()(int8_t& y, const int32_t& x, const int32_t& bias, const float& scaleAcc) const
{
float y_fp32 = ck::type_convert<float>(x + bias);
y_fp32 = scaleAcc * y_fp32;
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale2_ * y_fp32, -128.f, 127.f);
y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
float requantScale1_;
float requantScale2_;
__host__ __device__ constexpr void
operator()(int32_t& y, const int32_t& x, const int32_t& bias, const float& scaleAcc) const
{
// CAUSION - We might type_convert to int8 in threadwise copy
// eg. GridwiseGemmDlMultipleD_km_kn_mn
float y_fp32 = ck::type_convert<float>(x + bias);
y_fp32 = scaleAcc * y_fp32;
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int32_t>(y_fp32);
}
float scale_z_inv_;
Activation activationOp_;
};
......
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ bbe74503
......@@ -316,8 +316,36 @@ struct Sigmoid
y = 1 / (ck::type_convert<T>(1) + exp(-x));
};
};
int32_t divider_ = 1;
struct TanH
{
template <typename T>
__host__ __device__ void operator()(T& y, const T& x) const
{
static_assert(is_same<T, float>::value || is_same<T, double>::value ||
is_same<T, ck::half_t>::value,
"Data type is not supported by this operation!");
y = ck::math::tanh(x);
};
};
struct Swish
{
Swish(float beta = 1.0f) : beta_(beta) {}
template <typename T>
__host__ __device__ void operator()(T& y, const T& x) const
{
static_assert(is_same<T, float>::value || is_same<T, double>::value ||
is_same<T, ck::half_t>::value,
"Data type is not supported by this operation!");
y = x / (ck::type_convert<T>(1) + ck::math::exp(-beta_ * x));
};
float beta_ = 1.0f;
};
} // namespace element_wise
......
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View file @ bbe74503
......@@ -587,4 +587,52 @@ struct OffsettedBlockToCTileMap
index_t block_start_;
};
/**
* @brief Simple tile mapping which creates 3D grid of block of threads.
*
* @paragraph Description
* This Block-to-C-tile-map creates a 3D grid (n_blocks, m_blocks, z_blocks) of thread
* blocks. The first 2D are regular 2D tiles created by division of output GEMM
* dimenions by corresponding tile size. The third dimension (Z) is a k-split dimension,
* which denotes the number of blocks we use to divide work on GEMM K dimension onto.
*
* @tparam MPerBlock Output block tile size in M dimension.
* @tparam NPerBlock Output block tile size in N dimension.
*/
template <index_t MPerBlock, index_t NPerBlock>
struct BlockToCTileMap_3DGrid_KSplit
{
__host__ __device__ BlockToCTileMap_3DGrid_KSplit() = default;
__host__ __device__ constexpr auto
CalculateGridSize(index_t M, index_t N, index_t k_split) const
{
// Create 3D grid
const auto M0 = math::integer_divide_ceil(M, MPerBlock);
const auto N0 = math::integer_divide_ceil(N, NPerBlock);
return std::make_tuple(N0, M0, k_split);
}
template <typename TopIdx>
__device__ constexpr auto CalculateBottomIndex(const TopIdx&) const
{
return make_tuple(blockIdx.z, blockIdx.y, blockIdx.x);
}
template <typename CTileIdx, typename CTileDim>
__host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
const CTileDim& /* c_tile_dim */) const
{
return true; // always valid provided that user gets grid size from CalculateGridSize()
}
template <typename CGridDesc_M_N>
__host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
{
return true;
}
};
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
View file @ bbe74503
......@@ -66,7 +66,8 @@ __global__ void
const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
View file @ bbe74503
......@@ -185,8 +185,10 @@ struct GridwiseGemmDlMultipleD_km_kn_mn
return b_grid_desc_k0_n0_n1_k1;
}
// E desc for destination in blockwise copy
template <typename CGridDesc_M_N_>
__host__ __device__ static constexpr auto
MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(const CGridDesc_M_N& c_grid_desc_m_n)
MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(const CGridDesc_M_N_& c_grid_desc_m_n)
{
const auto M = c_grid_desc_m_n.GetLength(I0);
const auto N = c_grid_desc_m_n.GetLength(I1);
......@@ -238,19 +240,19 @@ struct GridwiseGemmDlMultipleD_km_kn_mn
using BGridDesc_K0_N0_N1_K1 = decltype(MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
using CGridDesc_M0_M10_M11_N0_N10_N11 =
decltype(MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
using Block2CTileMap = decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}));
using DsGridPointer = decltype(MakeDsGridPointer());
template <typename DsGridDesc_M0_M10_M11_N0_N10_N11,
bool HasMainKBlockLoop,
bool HasDoubleTailKBlockLoop>
bool HasDoubleTailKBlockLoop,
typename Block2CTileMap>
__device__ static void
Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
DsGridPointer p_ds_grid,
FloatC* __restrict__ p_c_grid,
FloatAB* __restrict__ p_shared_block,
void* __restrict__ p_shared_block,
const AElementwiseOperation&,
const BElementwiseOperation&,
const CDEElementwiseOperation& cde_element_op,
......@@ -399,8 +401,9 @@ struct GridwiseGemmDlMultipleD_km_kn_mn
constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
b_block_desc_k0_n0_n1_k1.GetElementSpaceSize(), max_lds_align);
FloatAB* p_a_block_double = p_shared_block;
FloatAB* p_b_block_double = p_shared_block + 2 * a_block_aligned_space_size;
FloatAB* p_a_block_double = static_cast<FloatAB*>(p_shared_block);
FloatAB* p_b_block_double =
static_cast<FloatAB*>(p_shared_block) + 2 * a_block_aligned_space_size;
// register allocation for output
auto c_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAcc>(
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
View file @ bbe74503
......@@ -431,6 +431,9 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
constexpr auto b_block_desc_k0perblock_nperblock_k1 =
GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
constexpr auto cshuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat();
constexpr auto max_lds_align = K1;
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
......@@ -439,8 +442,13 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize(), max_lds_align);
return (a_block_space_size_aligned * sizeof(ADataType) +
b_block_space_size_aligned * sizeof(BDataType));
constexpr auto c_block_space_size_aligned = math::integer_least_multiple(
cshuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat.GetElementSpaceSize(),
max_lds_align);
return math::max((a_block_space_size_aligned * sizeof(ADataType) +
b_block_space_size_aligned * sizeof(BDataType)),
c_block_space_size_aligned * sizeof(CShuffleDataType));
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
......@@ -497,6 +505,15 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
constexpr long_index_t TwoGB = (long_index_t{1} << 31);
if(!(a_grid_desc_k0_m_k1.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
b_grid_desc_k0_n_k1.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
{
return false;
}
return true;
}
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
View file @ bbe74503
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -92,6 +92,17 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
using GridwiseGemmPipe = remove_cvref_t<decltype(
GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
// denorm test fix, required to work around fp16 mfma issue
// we convert fp16->fp32->bf16 and execute bf16 mfma instruction
// when mfma if fixed, remove this section and update
// ABDataTypeAdjusted -> ABDataType throughout this file
#if CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
using ABDataTypeAdjusted =
conditional_t<is_same_v<ABDataType, ck::half_t>, ck::bhalf_t, ABDataType>;
#else
using ABDataTypeAdjusted = ABDataType;
#endif
__host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
{
// A matrix in LDS memory, dst of blockwise copy
......@@ -397,7 +408,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABDataType,
ABDataType,
ABDataTypeAdjusted,
decltype(a_grid_desc_ak0_m_ak1),
decltype(a_block_desc_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
......@@ -428,7 +439,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
ABDataType,
ABDataType,
ABDataTypeAdjusted,
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_block_desc_bk0_n_bk1),
BBlockTransferSrcAccessOrder,
......@@ -458,11 +469,11 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
// sanity check
constexpr index_t KPack =
math::max(math::lcm(AK1, BK1),
MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
MfmaSelector<ABDataTypeAdjusted, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
BlockSize,
ABDataType,
ABDataTypeAdjusted,
AccDataType,
decltype(a_block_desc_ak0_m_ak1),
decltype(b_block_desc_bk0_n_bk1),
......@@ -480,10 +491,11 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
static_cast<ABDataTypeAdjusted*>(p_shared),
a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
static_cast<ABDataTypeAdjusted*>(p_shared) + a_block_space_size_aligned,
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
View file @ bbe74503
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
View file @ bbe74503
......@@ -54,7 +54,8 @@ __global__ void
const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
View file @ bbe74503
......@@ -264,6 +264,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_wmma
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
constexpr long_index_t TwoGB = (long_index_t{1} << 31);
if(!(a_grid_desc_k0_m_k1.GetElementSpaceSize() * sizeof(FloatA) <= TwoGB &&
b_grid_desc_k0_n_k1.GetElementSpaceSize() * sizeof(FloatB) <= TwoGB))
{
return false;
}
return true;
}
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
View file @ bbe74503
......@@ -44,7 +44,8 @@ __global__ void
c_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
View file @ bbe74503
......@@ -57,7 +57,8 @@ __global__ void
const C0GridDescriptor_NBlock_NPerBlock c0_grid_desc_nblock_nperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
// TODO ANT: separate into MMA + Epilogue
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
View file @ bbe74503
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -165,16 +165,14 @@ __global__ void
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
__shared__ FloatAB p_shared_block[shared_block_size];
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
p_b_grid,
p_c_grid,
p_shared_block,
p_shared,
a_b_k0_m_k1_grid_desc,
b_b_k0_n_k1_grid_desc,
c_grid_desc_mblock_mperblock_nblock_nperblock,
......@@ -183,16 +181,16 @@ __global__ void
c_element_op,
c_block_cluster_adaptor);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_c_grid;
ignore = a_b_k0_m_k1_grid_desc;
ignore = b_b_k0_n_k1_grid_desc;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
ignore = c_block_cluster_adaptor;
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_c_grid;
ignore = a_b_k0_m_k1_grid_desc;
ignore = b_b_k0_n_k1_grid_desc;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
ignore = c_block_cluster_adaptor;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
......@@ -264,6 +262,16 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
using GridwiseGemmPipe = remove_cvref_t<decltype(
GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
// denorm test fix, required to work around fp16 mfma issue
// we convert fp16->fp32->bf16 and execute bf16 mfma instruction
// when mfma if fixed, remove this section and update
// FloatABAdjusted -> FloatAB throughout this file
#if CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
#else
using FloatABAdjusted = FloatAB;
#endif
// M0/M1/M1Padding
static constexpr auto M1PerBlock = Number<ABlockLdsM1PerBlock>{};
static constexpr auto M0PerBlock = Number<ABlockLdsM0PerBlock>{};
......@@ -605,7 +613,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
FloatAB* __restrict__ p_shared_block,
void* __restrict__ p_shared,
const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
......@@ -666,7 +674,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
FloatABAdjusted,
decltype(a_b_k0_m_k1_grid_desc),
decltype(a_b_k0_m_k1_block_desc),
ABlockTransferSrcAccessOrder,
......@@ -696,7 +704,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
FloatABAdjusted,
decltype(b_b_k0_n_k1_grid_desc),
decltype(b_b_k0_n_k1_block_desc),
BBlockTransferSrcAccessOrder,
......@@ -725,11 +733,11 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
// sanity check
constexpr index_t KPack =
math::max(K1, MfmaSelector<FloatAB, MPerXDL, NPerXDL>::selected_mfma.k_per_blk);
math::max(K1, MfmaSelector<FloatABAdjusted, MPerXDL, NPerXDL>::selected_mfma.k_per_blk);
auto blockwise_gemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatABAdjusted,
FloatAcc,
decltype(a_k0_m_k1_block_desc),
decltype(b_k0_n_k1_block_desc),
......@@ -745,16 +753,15 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
constexpr auto a_block_space_size =
math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
FloatAB* p_a_block = p_shared_block;
FloatAB* p_b_block = p_shared_block + a_block_space_size;
constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
static_cast<FloatABAdjusted*>(p_shared), a_k0_m_k1_block_desc.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
static_cast<FloatABAdjusted*>(p_shared) + a_block_space_size,
b_k0_n_k1_block_desc.GetElementSpaceSize());
// gridwise GEMM pipeline
const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
......@@ -798,8 +805,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
constexpr auto c_block_desc_mblock_mperblock_nblock_nperblock =
GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
void* p_shared = static_cast<void*>(p_shared_block);
auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatC*>(p_shared),
c_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
View file @ bbe74503
......@@ -44,7 +44,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
View file @ bbe74503
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -43,7 +43,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......@@ -58,16 +59,16 @@ __global__ void
c_element_op,
block_2_ctile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_c_grid;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1;
ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
ignore = block_2_ctile_map;
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_c_grid;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1;
ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
ignore = block_2_ctile_map;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
......@@ -131,6 +132,16 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
using GridwiseGemmPipe = remove_cvref_t<decltype(
GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
// denorm test fix, required to work around fp16 mfma issue
// we convert fp16->fp32->bf16 and execute bf16 mfma instruction
// when mfma if fixed, remove this section and update
// FloatABAdjusted -> FloatAB throughout this file
#if CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
#else
using FloatABAdjusted = FloatAB;
#endif
__host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
{
constexpr auto max_lds_align = K1;
......@@ -281,7 +292,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
using BlockwiseGemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatABAdjusted,
FloatAcc,
decltype(a_block_desc_k0_m_k1),
decltype(b_block_desc_k0_n_k1),
......@@ -367,7 +378,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
FloatABAdjusted,
decltype(a_grid_desc_k0_m_k1),
decltype(a_block_desc_k0_m_k1),
ABlockTransferSrcAccessOrder,
......@@ -398,7 +409,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
FloatABAdjusted,
decltype(b_grid_desc_k0_n_k1),
decltype(b_block_desc_k0_n_k1),
BBlockTransferSrcAccessOrder,
......@@ -428,7 +439,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
// sanity check
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
BlockSize,
FloatAB,
FloatABAdjusted,
FloatAcc,
decltype(a_block_desc_k0_m_k1),
decltype(b_block_desc_k0_n_k1),
......@@ -446,10 +457,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
static_cast<FloatABAdjusted*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
static_cast<FloatABAdjusted*>(p_shared) + a_block_space_size_aligned,
b_block_desc_k0_n_k1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
View file @ bbe74503
......@@ -42,7 +42,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
View file @ bbe74503
......@@ -15,63 +15,32 @@
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
namespace ck {
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename AGridDesc_B_K0_M_K1,
typename BGridDesc_B_K0_N_K1,
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename CBlockClusterAdaptor,
bool HasMainKBlockLoop>
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename Block2CTileMap>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_xdlops_v2r4r2(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const AGridDesc_B_K0_M_K1 a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1 b_b_k0_n_k1_grid_desc,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg,
const Block2CTileMap& b2c_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
__shared__ FloatAB p_shared_block[shared_block_size];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
p_b_grid,
p_c_grid,
static_cast<void*>(p_shared_block),
a_b_k0_m_k1_grid_desc,
b_b_k0_n_k1_grid_desc,
c_grid_desc_mblock_mperblock_nblock_nperblock,
a_element_op,
b_element_op,
c_element_op,
c_block_cluster_adaptor);
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
__shared__ uint8_t p_shared[shared_size];
GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>(
karg, static_cast<void*>(p_shared), b2c_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_c_grid;
ignore = a_b_k0_m_k1_grid_desc;
ignore = b_b_k0_n_k1_grid_desc;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
ignore = c_block_cluster_adaptor;
ignore = karg;
ignore = b2c_map;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
......@@ -79,13 +48,13 @@ template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_B_K0_M_K1,
typename BGridDesc_B_K0_N_K1,
typename CMNGridDesc,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
tensor_operation::device::GemmSpecialization GemmSpec,
index_t MPerBlock,
index_t NPerBlock,
index_t K0PerBlock,
......@@ -126,10 +95,238 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{};
static constexpr auto K1 = Number<K1Value>{};
static constexpr auto M01 = 1;
static constexpr auto N01 = 1;
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
struct Argument : public ck::tensor_operation::device::BaseArgument
{
const FloatAB* p_a_grid;
const FloatAB* p_b_grid;
FloatC* p_c_grid;
index_t M;
index_t N;
index_t K;
index_t StrideA;
index_t StrideB;
index_t StrideC;
index_t MPadded;
index_t NPadded;
index_t KPadded;
index_t K0;
index_t k_batch;
Argument(const FloatAB* p_a_grid_,
const FloatAB* p_b_grid_,
FloatC* p_c_grid_,
index_t M_,
index_t N_,
index_t K_,
index_t StrideA_,
index_t StrideB_,
index_t StrideC_,
index_t MPadded_,
index_t NPadded_,
index_t KPadded_,
index_t K0_,
index_t k_batch_)
: p_a_grid(p_a_grid_),
p_b_grid(p_b_grid_),
p_c_grid(p_c_grid_),
M(M_),
N(N_),
K(K_),
StrideA(StrideA_),
StrideB(StrideB_),
StrideC(StrideC_),
MPadded(MPadded_),
NPadded(NPadded_),
KPadded(KPadded_),
K0(K0_),
k_batch(k_batch_)
{
}
void Print() const
{
std::cout << "arg {"
<< "M:" << M << ", "
<< "N:" << N << ", "
<< "K:" << K << ", "
<< "SA:" << StrideA << ", "
<< "SB:" << StrideB << ", "
<< "SC:" << StrideC << ", "
<< "MP:" << MPadded << ", "
<< "NP:" << NPadded << ", "
<< "KP:" << KPadded << ", "
<< "K0:" << K0 << ", "
<< "KB:" << k_batch << "}" << std::endl;
}
};
__host__ __device__ static auto CalculateGridSize(const Argument& karg)
{
return std::make_tuple(math::integer_divide_ceil(karg.N, NPerBlock),
math::integer_divide_ceil(karg.M, MPerBlock),
karg.k_batch);
}
// prefer this to be called on host
__host__ __device__ static auto CalculateMPadded(index_t M)
{
return (M + MPerBlock - 1) / MPerBlock * MPerBlock;
}
__host__ __device__ static auto CalculateNPadded(index_t N)
{
return (N + NPerBlock - 1) / NPerBlock * NPerBlock;
}
__host__ __device__ static auto CalculateK0(index_t K, index_t K_Batch = 1)
{
// k_batch * k0 * k0_per_block * k1
auto K_t = K_Batch * K0PerBlock * K1;
return (K + K_t - 1) / K_t * K0PerBlock;
}
__host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
{
auto K0 = CalculateK0(K, K_Batch);
return K_Batch * K0 * K1;
}
__host__ __device__ static auto MakeAGridDescriptor_KBatch_K0_M_K1(index_t M,
index_t MPad,
index_t K,
index_t StrideA,
index_t KBatch,
index_t K0,
index_t KPad)
{
const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
}();
const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
a_grid_desc_m_k,
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>{}));
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)
{
// const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_right_pad_transform(M, MPad - M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else
{
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
}
__host__ __device__ static auto MakeBGridDescriptor_KBatch_K0_N_K1(index_t K,
index_t NPad,
index_t N,
index_t StrideB,
index_t KBatch,
index_t K0,
index_t KPad)
{
const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
}
}();
const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
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)
{
// const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else
{
return transform_tensor_descriptor(
b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
}
__host__ __device__ static auto
MakeCGridDescriptor_M_N(index_t M, index_t N, index_t MPad, index_t NPad, index_t StrideC)
{
const auto c_grid_desc_m_n = [&]() {
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));
}
}();
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
{
return transform_tensor_descriptor(c_grid_desc_m_n,
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
{
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
}
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
constexpr auto max_lds_align = K1;
......@@ -178,45 +375,68 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
c_block_size * sizeof(FloatC));
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
template <typename Block2CTileMap>
__host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
const CMNGridDesc& c_m_n_grid_desc,
const Block2CTileMap& block_2_ctile_map)
__host__ __device__ static constexpr bool CheckValidity(const Argument& karg)
{
static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
"wrong! K1 need to be known at compile-time");
static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
(NPerBlock % (NRepeat * NPerXDL)) == 0,
"Invalid tuning param!");
const auto M = a_b_k0_m_k1_grid_desc.GetLength(I2);
const auto N = b_b_k0_n_k1_grid_desc.GetLength(I2);
const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
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))
{
if(!(karg.M % MPerBlock == 0))
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))
{
if(!(karg.N % NPerBlock == 0))
return false;
}
if(!(M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
K0 == b_b_k0_n_k1_grid_desc.GetLength(I1) &&
K1 == a_b_k0_m_k1_grid_desc.GetLength(I3) &&
K1 == b_b_k0_n_k1_grid_desc.GetLength(I3) &&
KBatch == b_b_k0_n_k1_grid_desc.GetLength(I0)))
return false;
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
if(karg.K % ABlockTransferSrcScalarPerVector != 0)
return false;
}
else
{
if(karg.M % ABlockTransferSrcScalarPerVector != 0)
return false;
}
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
return false;
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
if(karg.N % BBlockTransferSrcScalarPerVector != 0)
return false;
}
else
{
if(karg.K % BBlockTransferSrcScalarPerVector != 0)
return false;
}
if(!block_2_ctile_map.CheckValidity(c_m_n_grid_desc))
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return false;
if(karg.N % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
return false;
}
else
{
if(karg.M % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
return false;
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return true;
}
__host__ __device__ static auto GetKPad(index_t K, index_t KBatch)
{
const index_t K0 = math::integer_divide_ceil(K, K1 * K0PerBlock * KBatch) * K0PerBlock;
const index_t KPad = KBatch * K0 * K1;
return KPad;
}
__host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
{
const bool has_main_k0_block_loop = K0 > K0PerBlock;
......@@ -224,8 +444,9 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
return has_main_k0_block_loop;
}
template <typename CGridDesc>
__host__ __device__ static constexpr auto
MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CMNGridDesc& c_m_n_grid_desc)
MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc& c_m_n_grid_desc)
{
const auto M = c_m_n_grid_desc.GetLength(I0);
const auto N = c_m_n_grid_desc.GetLength(I1);
......@@ -242,10 +463,11 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
}
// return block_id to C matrix tile idx (m0, n0) mapping
template <typename CGridDesc>
__host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
const CMNGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
const CGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
{
return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CMNGridDesc>(
return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc>(
c_m_n_grid_desc, 8, KBatch);
}
......@@ -262,24 +484,38 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
}
using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
decltype(MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CMNGridDesc{}));
using CBlockClusterAdaptor = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
// return block_id to C matrix tile idx (m0, n0, k_split) mapping
__host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap()
{
return BlockToCTileMap_3DGrid_KSplit<MPerBlock, NPerBlock>();
}
using CGridDesc_M_N = remove_cvref_t<decltype(MakeCGridDescriptor_M_N(1, 1, 1, 1, 1))>;
using DefaultBlock2CTileMap = remove_cvref_t<decltype(MakeDefaultBlock2CTileMap())>;
template <bool HasMainKBlockLoop>
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
template <bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename Block2CTileMap>
__device__ static void Run(const Argument& karg,
void* __restrict__ p_shared_block,
const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CElementwiseOperation& c_element_op,
const CBlockClusterAdaptor& c_block_cluster_adaptor)
const Block2CTileMap& block_2_ctile_map)
{
const FloatAB* p_a_grid = karg.p_a_grid;
const FloatAB* p_b_grid = karg.p_b_grid;
FloatC* p_c_grid = karg.p_c_grid;
const auto a_b_k0_m_k1_grid_desc = MakeAGridDescriptor_KBatch_K0_M_K1(
karg.M, karg.MPadded, karg.K, karg.StrideA, karg.k_batch, karg.K0, karg.KPadded);
const auto b_b_k0_n_k1_grid_desc = MakeBGridDescriptor_KBatch_K0_N_K1(
karg.K, karg.NPadded, karg.N, karg.StrideB, karg.k_batch, karg.K0, karg.KPadded);
const auto c_grid_desc_m_n =
MakeCGridDescriptor_M_N(karg.M, karg.N, karg.MPadded, karg.NPadded, karg.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
const AElementwiseOperation a_element_op = AElementwiseOperation{};
const BElementwiseOperation b_element_op = BElementwiseOperation{};
const CElementwiseOperation c_element_op = CElementwiseOperation{};
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
......@@ -287,28 +523,28 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
// divide block work by [M, N]
// divide block work by [KBatch, M, N]
const auto block_work_idx =
c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!c_block_cluster_adaptor.ValidCTileIndex(
make_tuple(block_work_idx[I1], block_work_idx[I2]),
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 k_batch_id = block_work_idx[I0];
const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I2]);
const index_t k_batch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
// 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_work_idx[I1] * MPerBlock);
__builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
__builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
// lds max alignment
constexpr auto max_lds_align = K1;
......@@ -444,7 +680,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register
// sanity check
#if 1
auto blockwise_gemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
......@@ -456,6 +692,20 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
MRepeat,
NRepeat,
K1>{};
#else
auto blockwise_gemm = BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<
BlockSize,
FloatAB,
FloatAcc,
decltype(a_k0_m_k1_block_desc),
decltype(b_k0_n_k1_block_desc),
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
K1>{};
#endif
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
......@@ -474,6 +724,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
#if 0
// preload data into LDS
{
a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
......@@ -519,6 +770,31 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
}
#else
// gridwise GEMM pipeline
const auto gridwise_gemm_pipeline =
GridwiseGemmPipeline_Selector<PipelineVersion::v2, 1, LoopScheduler::Default>();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_b_k0_m_k1_grid_desc.GetLength(I1) * a_b_k0_m_k1_grid_desc.GetLength(I3)) /
(K0PerBlock * K1));
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_b_k0_m_k1_grid_desc,
a_b_k0_m_k1_block_desc,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_b_k0_n_k1_grid_desc,
b_b_k0_n_k1_block_desc,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
blockwise_gemm,
c_thread_buf,
num_k_block_main_loop);
#endif
// output: register to global memory
{
......@@ -647,7 +923,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
{c_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0),
make_multi_index(block_m_id, 0, block_n_id, 0),
c_element_op};
constexpr auto mxdlperwave_forward_step =
......@@ -716,6 +992,48 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
});
}
}
template <typename Layout>
struct LStr
{
static std::string Get() { return ""; }
};
template <>
struct LStr<ck::tensor_layout::gemm::RowMajor>
{
static std::string Get() { return "R"; }
};
template <>
struct LStr<ck::tensor_layout::gemm::ColumnMajor>
{
static std::string Get() { return "C"; }
};
static std::string GetTypeString()
{
auto str = std::stringstream();
// clang-format off
str << "GemmXdlSplitKCShuffle_"
<< getGemmSpecializationString(GemmSpec) << "_"
<< std::string(ALayout::name)[0]
<< std::string(BLayout::name)[0]
<< std::string(CLayout::name)[0]
<< "_"
<< "B" << BlockSize << "_"
<< "Vec" << ABlockTransferSrcScalarPerVector << "x"
<< BBlockTransferSrcScalarPerVector << "x"
<< CBlockTransferScalarPerVector_NWaveNPerXDL << "_"
<< MPerBlock << "x"
<< NPerBlock << "x"
<< K0PerBlock << "x"
<< K1 ;
// clang-format on
return str.str();
}
};
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
View file @ bbe74503
......@@ -46,7 +46,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainK0BlockLoop>(
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
View file @ bbe74503
......@@ -49,7 +49,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
View file @ bbe74503
......@@ -53,7 +53,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(
......
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