1、英伟达GPU架构
Figure 1 shows a full GA100 GPU with 128 SMs. The A100 is based on GA100 and has 108 SMs.
SM是streaming multiprocessor的简写,4个处理单元组成一个SM,如Figure 2。
每个SM有64个INT32,64个FP32,32个FP64的CUDA core;每个SM还有4个Tensor Core。SM内共享L1缓存。
CUDA Core是用作通用计算的,Tensor Core是专门针对深度学习优化的,负责矩阵运算、混合精度运算。
Figure 1 GA100 Full GPU with 128 SMs. The A100 Tensor Core GPU has 108 SMs.
Figure 2. The GA100 streaming multiprocessor (SM)
Figure 3展示了NVDIA不同代GPU的特性。
Figure 3. 各代GPU架构
Tensor Core是从2017年的Volta架构开始演变的针对AI模型大量乘加运算的特殊处理单元。
在A100中,Tensor Core3.0能够支持所有的数据类型,包括FP16, BF16, TF32, FP64, INT8, INT4。提供了针对HPC用途的FP64双精度强大算力。
混合精度在底层硬件算子层面,使用半精度(FP16)作为输入和输出,使用全精度(FP32)进行中间结果计算从而不损失过多精度的技术。这个底层硬件层面其实指的就是Tensor Core,所以GPU上有Tensor Core是使用混合精度训练加速的必要条件。
可以说Tensor Core 是混合精度训练的底层硬件支持。
Tensor Core 的基本运算单元为 D = A*B + C,其中A、B、C、D 均为矩阵。每个 Tensor Core 能在一个时钟周期内完成 4*4 的 mma (matrix multiply accumalation) 运算,即一次矩阵乘法和一次矩阵加法。 (D = A*B + C是深度学习中最重要的算子,output = weight * input + bias)
Figure 4. 使用FP16做乘法,使用FP32做加法(避免大数吃小数)
2. MAGMA库
MAGMA的介绍:
MAGMA is intended for CUDA enabled NVIDIA GPUs and HIP enabled AMD GPUs.
It supports NVIDIA's Kepler, Maxwell, Pascal, Volta, Turing, Ampere, and Hopper
GPUs, and AMD's HIP GPUs.
Included are routines for the following algorithms:
* LU, QR, and Cholesky factorizations
* Hessenberg, bidiagonal, and tridiagonal reductions
* Linear solvers based on LU, QR, and Cholesky
* Eigenvalue and singular value (SVD) problem solvers
* Generalized Hermitian-definite eigenproblem solver
* Mixed-precision iterative refinement solvers based on LU, QR, and Cholesky
* MAGMA BLAS including gemm, gemv, symv, and trsm
* Batched MAGMA BLAS including gemm, gemv, herk, and trsm
* Batched MAGMA LAPACK including LU, inverse (getri), QR, and Cholesky factorizations
* MAGMA Sparse including CG, GMRES, BiCGSTAB, LOBPCG, iterative refinement,
preconditioners, sparse kernels (SpMV, SpMM), and support for CSR, ELL, and
SELL-P data formats
MAGMA的Dense Linear Algebra库是有混合精度算法(指Iterative Refinement,在MAGMA 2.5版本发布)。
比如下面两个接口magma_dsgesv_iteref_gpu和magma_dhgesv_iteref_gpu:
1- The FP32 to FP64 API magma_dsgesv_iteref_gpu, which is similar to the LAPACK dsgesv API. Here A, X, and B are FP64, the routine does the internal conversion and computation, and provides FP64 solution.
2- The FP16 to FP64 API magma_dhgesv_iteref_gpu, which is similar to the magma_dsgesv_gpu API, except it does use the tensor cores and performs computations in FP16. Here A, X, and B are FP64, the routine does the internal conversion and computation, and provides FP64 solution.
MAGMA 是在2018年时发布的MAGMA 2.5,支持了Tensor Core的计算,支持混合精度计算。
2.5.0 - Nov 16, 2018
* New routines: Magma is releasing the Nvidia Tensor Cores version
of its linear mixed-precision solver that is able to provide an
FP64 solution with up to 4X speedup. The release includes:
magma_dhgesv_iteref_gpu (FP64-FP16 solver with FP64 input and solution)
magma_dsgesv_iteref_gpu (FP64-FP32 solver with FP64 input and solution)
magma_hgetrf_gpu (mixed precision FP32-FP16 LU factorization)
magma_htgetrf_gpu (mixed precision FP32-FP16 LU factorization using Tensor Cores)
3. MAGMA Sparse
MAGMA Sparse支持稀疏线性系统的计算,但是solver的接口没有ds和cz开头的混合精度接口。
MAGMA Sparse中迭代法的算法,比如BICGSTAB,是在原始算法的基础上,做了很多算法重新设计以充分利用GPU的性能,并不是说用了混合精度就达到了这个加速目的。
下面左侧表示了原始BICGSTAB算法,右侧表示把这套算法简单粗暴地掉用CUBLAS算子实现,具体实现。这样做的话,并不能充分利用GPU的性能。需要作一些特殊的处理。
在MAGMA中的BICGSTAB主要做了下面的重设计,该算法称为merged-BICGSTAB:
- 将多个算术运算聚合(merge)到单个内核(kernel)中以减少GPU内存运输和CPU-GPU间的通信。
- 设计高效的dot kernel,允许多个点积的同时计算
性能对比时,指定双精度计算。得到如下的对比。
MAGMA Sparse是有内置的Iterative Refinement算法的,在magma_diterref、magma_siterref、magma_citerref和magma_ziterref等函数。
下面看一下magma_diterref的源码,可以发现算法实现并不复杂:
/*
-- MAGMA (version 2.8.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date March 2024
@author Hartwig Anzt
@generated from sparse/src/ziterref.cpp, normal z -> d, Thu Mar 28 12:29:00 2024
*/
#include "magmasparse_internal.h"
#define RTOLERANCE lapackf77_dlamch( "E" )
#define ATOLERANCE lapackf77_dlamch( "E" )
/**
Purpose
-------
Solves a system of linear equations
A * X = B
where A is a real symmetric N-by-N positive definite matrix A.
This is a GPU implementation of the Iterative Refinement method.
The inner solver is passed via the preconditioner argument.
Arguments
---------
@param[in]
A magma_d_matrix
input matrix A
@param[in]
b magma_d_matrix
RHS b
@param[in,out]
x magma_d_matrix*
solution approximation
@param[in,out]
solver_par magma_d_solver_par*
solver parameters
@param[in,out]
precond_par magma_d_preconditioner*
inner solver
@param[in]
queue magma_queue_t
Queue to execute in.
@ingroup magmasparse_dgesv
********************************************************************/
extern "C" magma_int_t
magma_diterref(
magma_d_matrix A, magma_d_matrix b, magma_d_matrix *x,
magma_d_solver_par *solver_par, magma_d_preconditioner *precond_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// some useful variables
double c_zero = MAGMA_D_ZERO;
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
// prepare solver feedback
solver_par->solver = Magma_ITERREF;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
magma_int_t dofs = A.num_rows*b.num_cols;
// solver variables
double nom, nom0;
// workspace
magma_d_matrix r={Magma_CSR}, z={Magma_CSR};
CHECK( magma_dvinit( &r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_dvinit( &z, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
double residual;
CHECK( magma_dresidual( A, b, *x, &residual, queue ));
solver_par->init_res = residual;
// solver setup
magma_dscal( dofs, c_zero, x->dval, 1, queue ); // x = 0
//CHECK( magma_dresidualvec( A, b, *x, &r, nom, queue));
magma_dcopy( dofs, b.dval, 1, r.dval, 1, queue ); // r = b
nom0 = magma_dnrm2( dofs, r.dval, 1, queue ); // nom0 = || r ||
nom = nom0 * nom0;
solver_par->init_res = nom0;
if( nom0 < solver_par->atol ||
nom0/solver_par->init_res < solver_par->rtol ){
solver_par->final_res = solver_par->init_res;
solver_par->iter_res = solver_par->init_res;
info = MAGMA_SUCCESS;
goto cleanup;
}
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = nom0;
solver_par->timing[0] = 0.0;
}
// start iteration
for( solver_par->numiter= 1; solver_par->numiter<solver_par->maxiter;
solver_par->numiter++ ) {
magma_dscal( dofs, MAGMA_D_MAKE(1./nom, 0.), r.dval, 1, queue ); // scale it
CHECK( magma_d_precond( A, r, &z, precond_par, queue )); // inner solver: A * z = r
magma_dscal( dofs, MAGMA_D_MAKE(nom, 0.), z.dval, 1, queue ); // scale it
magma_daxpy( dofs, c_one, z.dval, 1, x->dval, 1, queue ); // x = x + z
CHECK( magma_d_spmv( c_neg_one, A, *x, c_zero, r, queue )); // r = - A x
solver_par->spmv_count++;
magma_daxpy( dofs, c_one, b.dval, 1, r.dval, 1, queue ); // r = r + b
nom = magma_dnrm2( dofs, r.dval, 1, queue ); // nom = || r ||
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) nom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if( nom < solver_par->atol ||
nom/solver_par->init_res < solver_par->rtol ){
break;
}
}
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
CHECK( magma_dresidualvec( A, b, *x, &r, &residual, queue));
solver_par->final_res = residual;
solver_par->iter_res = nom;
if ( solver_par->numiter < solver_par->maxiter ) {
info = MAGMA_SUCCESS;
} else if ( solver_par->init_res > solver_par->final_res ) {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) nom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_SLOW_CONVERGENCE;
if( solver_par->iter_res < solver_par->atol ||
solver_par->iter_res/solver_par->init_res < solver_par->rtol ){
info = MAGMA_SUCCESS;
}
}
else {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) nom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_DIVERGENCE;
}
cleanup:
magma_dmfree(&r, queue );
magma_dmfree(&z, queue );
solver_par->info = info;
return info;
} /* magma_diterref */
MAGMA Sparse有一些混合精度的算子,详见magma-2.8.0\sparse\include\magmasparse_ds.h文件。
也有混合精度的SpMV的kernel如magma_dsgecsrmv_mixed_prec。
Purpose
-------
This routine computes y = alpha * A * x + beta * y on the GPU.
A is a matrix in mixed precision, i.e. the diagonal values are stored in
high precision, the offdiagonal values in low precision.
The input format is a CSR (val, row, col) in FloatComplex storing all
offdiagonal elements and an array containing the diagonal values in
DoubleComplex.
magma_dsgecsrmv_mixed_prec(
magma_trans_t transA,
magma_int_t m, magma_int_t n,
double alpha,
magmaDouble_ptr ddiagval,
magmaFloat_ptr doffdiagval,
magmaIndex_ptr drowptr,
magmaIndex_ptr dcolind,
magmaDouble_ptr dx,
double beta,
magmaDouble_ptr dy,
magma_queue_t queue )
4. Ginkgo库
Ginkgo是一个高性能的多核系统线性代数库,专注于求解稀疏线性系统。它是使用现代c++实现的(你至少需要一个c++ 14兼容的编译器来构建它),GPU内核在CUDA (NVIDIA设备),HIP (AMD设备)和SYCL/ dpc++ (Intel GPU)以及OpenMP (Intel/AMD/ARM multicore)实现。
Ginkgo的加速主要依赖混合精度算法,有下面三方面:
- 将存储精度与算术精度解耦。这能够加速内存绑定(memory-bound)算法的性能,可以补偿或容忍内存操作中的一些信息丢失。
- 允许不同精度格式的线性算子和向量之间的组合,而无需显式转换。
- 多种混合精度算法,比如SAI预处理、CB-GMRES等。
"Memory-bound algorithm"(内存绑定算法)是指在执行过程中主要受限于内存带宽或内存访问速度的算法。在计算机系统中,处理器(CPU)和内存(RAM)之间的数据传输速度通常是一个瓶颈。当算法的执行时间主要取决于内存访问的速度,而不是计算本身的速度时,就称该算法为内存绑定算法。
参考:
https://developer.nvidia.com/blog/nvidia-ampere-architecture-in-depth/
AI 工程师都应该知道的GPU工作原理,TensorCore_哔哩哔哩_bilibili
MAGMA的核心算法文献:https://icl.utk.edu/~tomov/MAGMA_Publications.bib
着重看下面的:
Magma Sparse中的BICGSTAB: Hartwig Anzt, 2014, Optimizing Krylov Subspace Solvers on Graphics Processing Units.
利用Tensor Core的mixed-precision iterative refinement solvers:Azzam Haidar, Stanimire Tomov, Jack Dongarra, and Nicholas J. Higham. 2018. Harnessing GPU tensor cores for fast FP16 arithmetic to speed up mixed-precision iterative refinement solvers.
Ginkgo的文献:Ginkgo: Citing Ginkgo
核心文献如下:
总体设计:Ginkgo: A Modern Linear Operator Algebra Framework for High Performance Computing
算法概览:Advances in Mixed Precision Algorithms 2021 Edition