CUDA Libraries and CUDA Fortran

Massimiliano Fatica
NVIDIA Corporation

1
NVIDIA CUDA Libraries
CUDA Toolkit includes several libraries:
— CUFFT: Fourier transforms
— CUBLAS: Dense Linear Algebra
— CUSPARSE : Sparse Linear Algebra
— LIBM: Standard C Math library
— CURAND: Pseudo-random and Quasi-random numbers
— NPP: Image and Signal Processing
— Thrust : Template Library

Several open source and commercial* libraries:

—MAGMA: Linear Algebra - OpenVidia: Computer Vision
—CULA Tools*: Linear Algebra - OpenCurrent: CFD
—CUSP: Sparse Linear Solvers ……..
—NAG*: Computational Finance

2
NVIDIA CUDA Libraries

Applications — CUFFT
— CUBLAS
3rd Party
— CUSPARSE
— Libm (math.h)
NVIDIA — CURAND
Libraries
— NPP
— Thrust
CUDA C/Fortran — CUSP

3
CUFFT Library

CUFFT is a GPU based Fast Fourier Transform library

4
 CUFFT Library Features
Cooley-Tukey

 Algorithms based on Cooley-Tukey (n = 2a · 3b · 5c · 7d) and

Bluestein N1 N2

 Simple interface similar to FFTW N2 T,T N1

 1D, 2D and 3D transforms of complex and real data

 Row-major order (C-order) for 2D and 3D data
 Single precision (SP) and Double precision (DP) transforms
 In-place and out-of-place transforms
 1D transform sizes up to 128 million elements
 Batch execution for doing multiple transforms
 Streamed asynchronous execution
 Non normalized output: IFFT(FFT(A))=len(A)*A

5
CUFFT in 4 easy steps

Step 1 – Allocate space on GPU memory

Step 2 – Create plan specifying transform configuration like the size

and type (real, complex, 1D, 2D and so on).

Step 3 – Execute the plan as many times as required, providing the

pointer to the GPU data created in Step 1.

Step 4 – Destroy plan, free GPU memory

6
Code example:
#define NX 256
#define NY 128

cufftHandle plan;
cufftComplex *idata, *odata;
cudaMalloc((void**)&idata, sizeof(cufftComplex)*NX*NY);
cudaMalloc((void**)&odata, sizeof(cufftComplex)*NX*NY);
…
/* Create a 2D FFT plan. */
cufftPlan2d(&plan, NX,NY, CUFFT_C2C);

/* Use the CUFFT plan to transform the signal out of place. */

cufftExecC2C(plan, idata, odata, CUFFT_FORWARD);

/* Inverse transform the signal in place.

Different pointers to input and output arrays implies out of place transformation */
cufftExecC2C(plan, odata, odata, CUFFT_INVERSE);
….
/* Destroy the CUFFT plan. */
cufftDestroy(plan);

cudaFree(idata), cudaFree(odata);

7
CUBLAS Library
 Implementation of BLAS (Basic Linear Algebra Subprograms)
 Self-contained at the API level
 Supports all the BLAS functions
— Level1 (vector,vector): O(N)
 AXPY : y = alpha.x + y
 DOT : dot = x.y

— Level 2( matrix,vector): O(N2)

 Vector multiplication by a General Matrix : GEMV
 Triangular solver : TRSV

— Level3(matrix,matrix): O(N3)
 General Matrix Multiplication : GEMM
 Triangular Solver : TRSM

 Following BLAS convention, CUBLAS uses column-major storage

 Function naming convention: cublas + BLAS name

8
Using CUBLAS
Interface to CUBLAS library is in cublas.h
Function naming convention
cublas + BLAS name
Eg., cublasSGEMM
Error handling
CUBLAS core functions do not return error
CUBLAS provides function to retrieve last error recorded
CUBLAS helper functions do return error
Helper functions:
Memory allocation, data transfer

9
Calling CUBLAS from C

10
Calling CUBLAS from FORTRAN
Two interfaces:

Thunking
Allows interfacing to existing applications without any changes
During each call, the wrappers allocate GPU memory, copy source data from CPU memory space to GPU
memory space, call CUBLAS, and finally copy back the results to CPU memory space and deallocate the
GPGPU memory
Intended for light testing due to call overhead

Non-Thunking (default)
Intended for production code
Substitute device pointers for vector and matrix arguments in all BLAS functions
Existing applications need to be modified slightly to allocate and deallocate data structures in GPGPU
memory space (using CUBLAS_ALLOC and CUBLAS_FREE) and to copy data between GPU and CPU
memory spaces (using CUBLAS_SET_VECTOR, CUBLAS_GET_VECTOR, CUBLAS_SET_MATRIX, and
CUBLAS_GET_MATRIX)

11
SGEMM example (THUNKING)
program example_sgemm
! Define 3 single precision matrices A, B, C
real, dimension(:,:),allocatable:: A(:,:),B(:,:),C(:,:)
integer:: n=16
allocate (A(n,n),B(n,n),C(n,n))
! Initialize A, B and C
…
#ifdef CUBLAS
! Call SGEMM in CUBLAS library using THUNKING interface (library takes care of
! memory allocation on device and data movement)
call cublas_SGEMM('n','n', n,n,n,1.,A,n,B,n,1.,C,n)
#else
! Call SGEMM in host BLAS library
call SGEMM ('n','n',m1,m1,m1,alpha,A,m1,B,m1,beta,C,m1)
#endif
print *,c(n,n)
end program example_sgemm

To use the host BLAS routine:

g95 –O3 code.f90 –L/usr/local/lib64 -lblas

To use the CUBLAS routine (fortran_thunking.c is included in the toolkit /usr/local/cuda/src):

nvcc -O3 -c fortran_thunking.c
g95 -O3 -DCUBLAS code.f90 fortran_thunking .o -L/usr/local/cuda/lib64 -lcudart -lcublas

12
SGEMM example (NON-THUNKING)
program example_sgemm
real, dimension(:,:),allocatable:: A(:,:),B(:,:),C(:,:)
integer*8:: devPtrA, devPtrB, devPtrC
integer:: n=16, size_of_real=16
allocate (A(n,n),B(n,n),C(n,n))
call cublas_Alloc(n*n,size_of_real, devPtrA)
call cublas_Alloc(n*n,size_of_real, devPtrB)
call cublas_Alloc(n*n,size_of_real, devPtrC)
! Initialize A, B and C
…
! Copy data to GPU
call cublas_Set_Matrix(n,n,size_of_real,A,n,devPtrA,n)
call cublas_Set_Matrix(n,n,size_of_real,B,n,devPtrB,n)
call cublas_Set_Matrix(n,n,size_of_real,C,n,devPtrC,n)
! Call SGEMM in CUBLAS library
call cublas_SGEMM('n','n', n,n,n,1.,devPtrA,n,devPtrB,n,1.,devPtrC,n)
! Copy data from GPU
call cublas_Get_Matrix(n,n,size_of_real,devPtrC,n,C,n)
print *,c(n,n)
call cublas_Free(devPtrA)
call cublas_Free(devPtrB)
call cublas_Free(devPtrC)
end program example_sgemm

To use the CUBLAS routine (fortran.c is included in the toolkit /usr/local/cuda/src):

nvcc -O3 -c fortran.c
g95 -O3 code.f90 fortran.o -L/usr/local/cuda/lib64 -lcudart -lcublas

13
Using CPU and GPU concurrently

DGEMM(A,B,C) = DGEMM(A,B1,C1) U DGEMM(A,B2,C2)

(GPU) (CPU)
The idea can be extended to multi-GPU configuration and to handle huge matrices

Find the optimal split, knowing the relative performances of the GPU and CPU cores on DGEMM

14
 Overlap DGEMM on CPU and GPU
// Copy A from CPU memory to GPU memory devA
status = cublasSetMatrix (m, k , sizeof(A[0]), A, lda, devA, m_gpu);
// Copy B1 from CPU memory to GPU memory devB
status = cublasSetMatrix (k ,n_gpu, sizeof(B[0]), B, ldb, devB, k_gpu);
// Copy C1 from CPU memory to GPU memory devC
status = cublasSetMatrix (m, n_gpu, sizeof(C[0]), C, ldc, devC, m_gpu);

// Perform DGEMM(devA,devB,devC) on GPU

// Control immediately return to CPU
cublasDgemm('n', 'n', m, n_gpu, k, alpha, devA, m,devB, k, beta, devC, m);

// Perform DGEMM(A,B2,C2) on CPU

dgemm('n','n',m,n_cpu,k, alpha, A, lda,B+ldb*n_gpu, ldb, beta,C+ldc*n_gpu, ldc);

// Copy devC from GPU memory to CPU memory C1

status = cublasGetMatrix (m, n, sizeof(C[0]), devC, m, C, *ldc);

Using CUBLAS, it is very easy to express the workflow in the diagram

15
CUDA Libm features

High performance and high accuracy implementation:

• C99 compatible math library, plus extras
• Basic ops: x+y, x*y, x/y, 1/x, sqrt(x), FMA (IEEE-754 accurate in single, double)
• Exponentials: exp, exp2, log, log2, log10, ...
• Trigonometry: sin, cos, tan, asin, acos, atan2, sinh, cosh, asinh, acosh, ...
• Special functions: lgamma, tgamma, erf, erfc
• Utility: fmod, remquo, modf, trunc, round, ceil, floor, fabs, ...
• Extras: rsqrt, rcbrt, exp10, sinpi, sincos, erfinv, erfcinv, ...

16
CURAND Library

Library for generating random numbers

Features:
XORWOW pseudo-random generator
Sobol’ quasi-random number generators
Host API for generating random numbers in bulk
Inline implementation allows use inside GPU functions/kernels
Single- and double-precision, uniform, normal and log-normal distributions

17
 CURAND use
1. Create a generator:
curandCreateGenerator()
2. Set a seed:
curandSetPseudoRandomGeneratorSeed()
3. Generate the data from a distribution:
curandGenerateUniform()/(curandGenerateUniformDouble(): Uniform
curandGenerateNormal()/cuRandGenerateNormalDouble(): Gaussian
curandGenerateLogNormal/curandGenerateLogNormalDouble(): Log-Normal
4. Destroy the generator:
curandDestroyGenerator()

18
Example CURAND Program: Host API

19
Example CURAND Program: Run on CPU

20
 NVIDIA Performance Primitives (NPP)

– C library of functions (primitives)

– well optimized
– low level API:
– easy integration into existing code
– algorithmic building blocks
– actual operations execute on CUDA GPUs
– Approximately 350 image processing
functions
– Approximately 100 signal processing
functions

21
Image Processing Primitives

• Data exchange & initialization • Filter Functions

– Set, Convert, CopyConstBorder, Copy, – FilterBox, Row, Column, Max, Min, Dilate,
Transpose, SwapChannels Erode, SumWindowColumn/Row

• Arithmetic & Logical Ops • Geometry Transforms

– Add, Sub, Mul, Div, AbsDiff – Resize , Mirror, WarpAffine/Back/Quad,
WarpPerspective/Back/Quad
• Threshold & Compare Ops
– Threshold, Compare • Statistics
– Mean, StdDev, NormDiff, MinMax,
• Color Conversion Histogram, SqrIntegral, RectStdDev
– RGB To YCbCr (& vice versa), ColorTwist,
LUT_Linear • Segmentation
– Graph Cut

22
Thrust
 A template library for CUDA
— Mimics the C++ STL

 Containers
— Manage memory on host and device: thrust::host_vector<T>, thrust:device_vector<T>
— Help avoid common errors
 Iterators
— Know where data lives
— Define ranges: d_vec.begin()
 Algorithms
— Sorting, reduction, scan, etc: thrust::sort()
— Algorithms act on ranges and support general types and operators

23
Thrust Example
#include <thrust/host_vector.h>
#include <thrust/device_vector.h>
#include <thrust/sort.h>
#include <cstdlib.h>

int main(void)
{
// generate 32M random numbers on the host
thrust::host_vector<int> h_vec(32 << 20);
thrust::generate(h_vec.begin(), h_vec.end(), rand);

// transfer data to the device

thrust::device_vector<int> d_vec = h_vec;

// sort data on the device (846M keys per sec on GeForce GTX 480)
thrust::sort(d_vec.begin(), d_vec.end());

// transfer data back to host

thrust::copy(d_vec.begin(), d_vec.end(), h_vec.begin());

return 0;
}

24
 Algorithms
 Elementwise operations
— for_each, transform, gather, scatter …
 Reductions
— reduce, inner_product, reduce_by_key …
 Prefix-Sums
— inclusive_scan, inclusive_scan_by_key …
 Sorting
— sort, stable_sort, sort_by_key …

25
Interoperability (from Thrust to C/CUDA)

 Convert iterators to raw pointers

// allocate device vector

thrust::device_vector<int> d_vec(4);

// obtain raw pointer to device vector’s memory

int * ptr = thrust::raw_pointer_cast(&d_vec[0]);

// use ptr in a CUDA C kernel

my_kernel<<< N / 256, 256 >>>(N, ptr);

// Note: ptr cannot be dereferenced on the host!

26
 Interoperability (from C/CUDA to Thrust)
 Wrap raw pointers with device_ptr

// raw pointer to device memory

int * raw_ptr;
cudaMalloc((void **) &raw_ptr, N * sizeof(int));

// wrap raw pointer with a device_ptr

device_ptr<int> dev_ptr(raw_ptr);

// use device_ptr in thrust algorithms

fill(dev_ptr, dev_ptr + N, (int) 0);

// access device memory through device_ptr

dev_ptr[0] = 1;

// free memory
cudaFree(raw_ptr);

27
Introduction to CUDA Fortran

28
Introduction
• CUDA is a scalable model for parallel computing

• CUDA Fortran is the Fortran analog to CUDA C

– Program has host and device code similar to CUDA C
– Host code is based on the runtime API
– Fortran language extensions to simplify data management

• Co-defined by NVIDIA and PGI, implemented in the PGI Fortran

compiler

29
CUDA Programming
• Heterogeneous programming model
– CPU and GPU are separate devices with separate memory spaces
– Host code runs on the CPU
• Handles data management for both host and device
• Launches kernels which are subroutines executed on the GPU
– Device code runs on the GPU
• Executed by many GPU threads in parallel
– Allows for incremental development

30
Heterogeneous Programming
• Host = CPU and its memory
• Device = GPU and its memory
Host Device
• Typical code progression CPU GPU
– Allocate memory on host and device
– Transfer data from host to device PCIe
DRAM DRAM
– Execute kernel (device computation)
– Transfer result from device to host
– Deallocate memory

31
Data Transfers
program copyData
use cudafor
implicit none Host Device
integer, parameter :: n = 256
real :: a(n), b(n) CPU GPU
real, device :: a_d(n), b_d(n)
DRAM DRAM
a = 1.0
a_d = a a a_d
b_d = a_d
b = b_d b b_d
if (all(a == b)) &
write(*,*) 'Test Passed'
end program copyData

32
Data Transfers
program copyData
use cudafor
implicit none Host Device
integer, parameter :: n = 256
real :: a(n), b(n) CPU GPU
real, device :: a_d(n), b_d(n)
DRAM DRAM
a = 1.0
a_d = a a a_d
b_d = a_d
b = b_d b b_d
if (all(a == b)) &
write(*,*) 'Test Passed'
end program copyData

33
Data Transfers
program copyData
use cudafor
implicit none Host Device
integer, parameter :: n = 256
real :: a(n), b(n) CPU GPU
real, device :: a_d(n), b_d(n)
DRAM DRAM
a = 1.0 PCIe
a_d = a a a_d
b_d = a_d
b = b_d b b_d
if (all(a == b)) &
write(*,*) 'Test Passed'
end program copyData

34
Data Transfers
program copyData
use cudafor
implicit none Host Device
integer, parameter :: n = 256
real :: a(n), b(n) CPU GPU
real, device :: a_d(n), b_d(n)
DRAM DRAM
a = 1.0
a_d = a a a_d
b_d = a_d
b = b_d b b_d
if (all(a == b)) &
write(*,*) 'Test Passed'
end program copyData

35
Data Transfers
program copyData
use cudafor
implicit none Host Device
integer, parameter :: n = 256
real :: a(n), b(n) CPU GPU
real, device :: a_d(n), b_d(n)
DRAM DRAM
a = 1.0
a_d = a a a_d
b_d = a_d PCIe
b = b_d b b_d
if (all(a == b)) &
write(*,*) 'Test Passed'
end program copyData

36
F90 Array Increment
module simpleOps_m program incTest
contains use simpleOps_m
subroutine inc(a, b) implicit none
implicit none integer :: b, n = 256
integer :: a(:) integer, allocatable:: a(:)
integer :: b
integer :: i, n allocate (a(n))
a = 1 ! array assignment
n = size(a) b = 3
do i = 1, n call inc(a, b)
a(i) = a(i)+b
enddo if (all(a == 4)) &
write(*,*) 'Test Passed'
end subroutine inc
end module simpleOps_m deallocate(a)
end program incTest

37
CUDA Fortran - Host Code
F90 CUDA Fortran
program incTest program incTest
use cudafor
use simpleOps_m use simpleOps_m
implicit none implicit none
integer :: b, n = 256 integer :: b, n = 256
integer, allocatable:: a(:) integer, allocatable :: a(:)
integer, allocatable, device :: a_d(:)
allocate (a(n)) allocate (a(n),a_d(n))
a = 1 ! array assignment a = 1
b = 3 b = 3
a_d = a
call inc(a, b) call inc<<<1,n>>>(a_d, b)
a = a_d
if (all(a == 4)) & if (all(a == 4)) &
write(*,*) 'Test Passed' write(*,*) 'Test Passed'
deallocate(a)
deallocate (a,a_d)
end program incTest end program incTest

38
CUDA Fortran - Device Code
F90 CUDA Fortran
module simpleOps_m module simpleOps_m
contains contains
subroutine inc(a, b) attributes(global) subroutine inc(a, b)
implicit none implicit none
integer :: a(:) integer :: a(:)
integer :: b integer, value :: b
integer :: i, n integer :: i
n = size(a) i = threadIdx%x
do i = 1, n a(i) = a(i)+b
a(i) = a(i)+b
enddo
end subroutine inc end subroutine inc
end module simpleOps_m end module simpleOps_m

39
Extending to Larger Arrays
• Previous example works with small arrays

call inc<<<1,n>>>(a_d,b)

– Limit of n=1024 (Fermi) or n=512 (pre-Fermi)

• For larger arrays, change the first execution parameter

(<<<1,n>>>)

40
Execution Model
Software Hardware
Threads are executed by thread
Thread processors
Thread Processor

Thread blocks are executed on

multiprocessors

Thread blocks do not migrate

Thread Several concurrent thread blocks can reside

Block Multiprocessor on a multiprocessor

A kernel is launched on a device as a

... grid of thread blocks

Grid Device
41
Execution Configuration
• Execution configuration specified in host code

call inc<<<blocksPerGrid, threadsPerBlock>>>(a_d,b)

• Previous example used a single thread block

call inc<<<1,n>>>(a_d,b)

• Multiple thread blocks

tPB = 256
call inc<<<ceiling(real(n)/tPB),tPB>>>(a_d,b)

42
Mapping Arrays to Thread Blocks
• call inc<<<3,4>>>(a_d, b)

blockDim%x = 4

blockIdx%x 1 2 3

threadIdx%x 1 2 3 4 1 2 3 4 1 2 3 4

(blockIdx%x-1)*blockDim%x
1 2 3 4 5 6 7 8 9 10 11 12
+ threadIdx%x

43
Large Array - Host Code
program incTest
use cudafor
use simpleOps_m
implicit none
integer, parameter :: n = 1024*1024
integer, parameter :: tPB = 256
integer :: a(n), b
integer, device :: a_d(n)
a = 1
b = 3
a_d = a
call inc<<<ceiling(real(n)/tPB),tPB>>>(a_d, b)
a = a_d
if (all(a == 4)) then
write(*,*) 'Success'
endif
end program incTest

44
Built-in Variables for Device Code
• Predefined variables in device subroutines
– Grid and block dimensions - gridDim, blockDim
– Block and thread indices - blockIdx, threadIdx
– Of type dim3
type (dim3)
integer (kind=4) :: x, y, z
end type

– blockIdx and threadIdx fields have unit offset

1 <= blockIdx%x <= gridDim%x

45
Large Array - Device Code

module simpleOps_m
contains
attributes(global) subroutine inc(a, b)
implicit none
integer :: a(:)
integer, value :: b
integer :: i, n
i = (blockIdx%x-1)*blockDim%x + threadIdx%x
n = size(a)
if (i <= n) a(i) = a(i)+ b
end subroutine inc
end module simpleOps_m

46
Multidimensional Example - Host
• Execution Configuration

call inc<<<blocksPerGrid, threadsPerBlock>>>(a_d,b)

– Grid dimensions in blocks (blocksPerGrid) and block dimensions in

threads (threadsPerBlock) can be either integer or dim3

type (dim3)
integer (kind=4) :: x, y, z
end type

47
2D Example - Host Code
program incTest
use cudafor
use simpleOps_m
implicit none
integer, parameter :: nx=1024, ny=512
real :: a(nx,ny), b
real, device :: a_d(nx,ny)
type(dim3) :: grid, tBlock

a = 1; b = 3

tBlock = dim3(32,8,1)
grid = dim3(ceiling(real(nx)/tBlock%x), ceiling(real(ny)/tBlock%y), 1)
a_d = a
call inc<<<grid,tBlock>>>(a_d, b)
a = a_d

write(*,*) 'Max error: ', maxval(abs(a-4))

end program incTest

48
2D Example - Device Code
module simpleOps_m
contains
attributes(global) subroutine inc(a, b)
implicit none
real :: a(:,:)
real, value :: b
integer :: i, j

i = (blockIdx%x-1)*blockDim%x + threadIdx%x
j = (blockIdx%y-1)*blockDim%y + threadIdx%y

if (i<=size(a,1) .and. j<=size(a,2)) &

a(i,j) = a(i,j) + b

end subroutine inc

end module simpleOps_m

49
Kernel Loop Directives (CUF Kernels)
• Automatic kernel generation and invocation of host code
region (arrays in loops must reside on GPU)
program incTest
use cudafor
implicit none
integer, parameter :: n = 256
integer :: a(n), b
integer, device :: a_d(n)

a = 1; b = 3; a_d = a

!$cuf kernel <<<*,*>>>

do i = 1, n
a_d(i) = a_d(i)+b
enddo

a = a_d
if (all(a == 4)) write(*,*) 'Test Passed'
end program incTest

50
Kernel Loop Directives (CUF Kernels)
• Multidimensional arrays
!$cuf kernel do(2) <<< *,* >>>
do j = 1, ny
do i = 1, nx
a_d(i,j) = b_d(i,j) + c_d(i,j)
enddo
enddo

• Can specify parts of execution parameter

!$cuf kernel do(2) <<<(*,*),(32,4)>>>

51
Reduction using CUF Kernels
• Compiler recognizes use of scalar reduction and generates one
result

rsum = 0.0
!$cuf kernel do <<<*,*>>>
do i = 1, nx
rsum = rsum + a_d(i)
enddo

52
Compute Capabilities
Architecture Tesla Fermi Kepler

Compute capabilities 1.0 1.1 1.2 1.3 2.0 2.1 3.0

Double precision

3D grids

Max # threads per block 512 1024

Shared memory per MP 16Kb 48Kb 48Kb

(16/48,48/16) +(32/32)

..... ..... ... ..

• All these values are returned by cudaGetDeviceProperties

• Target GPU can be specified with -Mcuda=ccx.x
53
Compilation
• pgfortran - PGI’s Fortran compiler
– All source code with .cuf or .CUF is compiled as CUDA Fortran enabled
automatically
– Flag to target architecture (eg. -Mcuda=cc20)
– -Mcuda=emu specifies emulation mode
– Flag to target CUDA Toolkit version (eg. -Mcuda=cuda4.0)
– -Mcuda=fastmath enables faster intrinsics (__sinf())
– -Mcuda=nofma turns of fused multiply-add
– -Mcuda=maxregcount:<n> limits register usage per thread
– -Mcuda=ptxinfo prints memory usage per kernel
• pgfortran -Mcuda -help for a full list

54
Host-Device Transfers
• Host-device bandwidth is much lower than bandwidth within
device
– 8 GB/s peak (PCIe x16 Gen 2) vs. 177 GB/s peak (Tesla M2090)

• Minimize number of transfers

– Intermediate data can be allocated, used, and deallocated without
copying to host memory
– Sometimes better to do low parallelism operations on the GPU if it
avoids transfers to and from host

55
Page-Locked Data Transfers

Pageable Data Transfer Page-locked Data Transfer

Device Device
DRAM DRAM

Host Host
Pageable Pinned Pinned
memory buffer memory

56
Page-Locked Data Transfers
• Page-locked or pinned host memory by declaration
– Designated by pinned variable attribute
– Must be allocatable

real, device :: a_d(N)

real, pinned, allocatable :: a(:)
allocate(a(N), STAT=istat, PINNED=pinnedFlag)
...
a_d = a

– Tesla M2050/Nehalem
• Pageable: ~3.5 GB/s
• Pinned: ~6 GB/s

57
Overlapping Transfers and Computation
• Kernel launches are asynchronous, normal memory copies are
blocking
a_d = a ! blocks on host until transfer completes
call inc<<<g,b>>>(a_d, b) ! Control returns immediately to CPU
a = a_d ! starts only after kernel completes

• Asynchronous and Stream APIs allow overlap of transfers with

computation
• A stream is a sequence of operations that execute in order on
the GPU
– Operations in different (non-default) streams can be interleaved
– Stream ID used as arguments to async transfers and kernel launches

58
Asynchronous Data Transfers
• Asynchronous host-device transfers return control immediately
to CPU
– cudaMemcpyAsync(dst, src, nElements, stream)
– Requires pinned host memory

• Overlapping data transfer with CPU computation

– default stream = 0

istat = cudaMemcpyAsync(a_d, a_h, N, 0)

call kernel<<<grid, block>>>(a_d)
call cpuFunction(b)
overlapped

59
Overlapping Transfers and Kernels
• Requires:
– Pinned host memory
– Kernel and transfer to use different non-zero streams

integer (kind=cuda_stream_kind) :: stream1, stream2

...
istat = cudaStreamCreate(stream1)
istat = cudaStreamCreate(stream2)
istat = cudaMemcpyAsync(a_d, a_h, N, stream1)
call kernel<<<grid, block, 0, stream2>>>(b_d)
overlapped

60
GPU/CPU Synchronization
• cudaDeviceSynchronize()
– Blocks until all previously issued operations on the GPU complete

• cudaStreamSynchronize(stream)
– Blocks until all previously issued operations to stream complete

• cudaStreamQuery(stream)
– Indicated whether stream is idle
– Does not block CPU code

61
GPU/CPU Synchronization
• Stream-based using CUDA events
– Events can be inserted into streams

type (cudaEvent) :: event

...
istat = cudaEventRecord(event, stream1)

– Event is recorded when the GPU reaches it in the stream

• Recorded = assigned a time stamp
• Useful for timing code

– cudaEventSynchronize(event)
• Blocks CPU until event is recorded

62
Shared Memory
• On-chip
• All threads in a block have access to same shared memory
• Used to reduce multiple loads of device data
• Used to accommodate coalescing

DRAM

Shared Shared

63
Matrix Transpose
attributes(global) subroutine transposeNaive(odata, idata)
real, intent(out) :: odata(ny,nx)
real, intent(in) :: idata(nx,ny)
integer :: x, y

x = (blockIdx%x-1) * blockDim%x + threadIdx%x

y = (blockIdx%y-1) * blockDim%y + threadIdx%y
odata(y,x) = idata(x,y)
end subroutine transposeNaive
idata odata

64
Matrix Transpose - Shared Memory
attributes(global) subroutine transposeCoalesced(odata, idata)
real, intent(out) :: odata(ny,nx)
real, intent(in) :: idata(nx,ny)
real, shared :: tile(TILE_DIM, TILE_DIM)
integer :: x, y

x = (blockIdx%x-1)*blockDim%x + threadIdx%x
y = (blockIdx%y-1)*blockDim%y + threadIdx%y
tile(threadIdx%x, threadIdx%y) = idata(x,y)

call syncthreads()

x = (blockIdx%y-1)*blockDim%y + threadIdx%x
y = (blockIdx%x-1)*blockDim%x + threadIdx%y idata odata
odata(x,y) = tile(threadIdx%y, threadIdx%x)
end subroutine transposeCoalesced tile

65
Calling CUBLAS from CUDA Fortran
• Module which defines interfaces to CUBLAS from CUDA Fortran
– use cublas
• Interfaces in three forms
– Overloaded BLAS interfaces that take device array arguments
• call saxpy(n, a_d, x_d, incx, y_d, incy)
– Legacy CUBLAS interfaces
• call cublasSaxpy(n, a_d, x_d, incx, y_d, incy)
– Multi-GPU version (CUDA 4.0) that utilizes a handle h
• istat = cublasSaxpy_v2(h, n, a_d, x_d, incx, y_d, incy)
• Mixing the three forms is allowed

66
Calling CUBLAS from CUDA Fortran
program cublasTest
use cublas
implicit none

real, allocatable :: a(:,:),b(:,:),c(:,:)

real, device, allocatable :: a_d(:,:),b_d(:,:),c_d(:,:)
integer :: k=4, m=4, n=4
real :: alpha=1.0, beta=2.0, maxError

allocate(a(m,k), b(k,n), c(m,n), a_d(m,k), b_d(k,n), c_d(m,n))

a = 1; a_d = a
b = 2; b_d = b
c = 3; c_d = c

call cublasSgemm('N','N',m,n,k,alpha,a_d,m,b_d,k,beta,c_d,m) ! or sgemm(..)

c=c_d
write(*,*) 'Maximum error: ', maxval(abs(c-14.0))

deallocate (a,b,c,a_d,b_d,c_d)

end program cublasTest

67
 Calling Thrust from CUDA Fortran
C wrapper for Thrust: csort.cu

#include <thrust/device_vector.h> //Sort for single precision arrays

#include <thrust/device_vector.h> void sort_float_wrapper( float *data, int N)
#include <thrust/sort.h> {
thrust::device_ptr<float> dev_ptr(data);
extern "C" { thrust::sort(dev_ptr, dev_ptr+N);
//Sort for integer arrays }
void sort_int_wrapper( int *data, int N)
{ //Sort for double precision arrays
// Wrap raw pointer with a device_ptr void sort_double_wrapper(double *data, int N)
thrust::device_ptr<int> dev_ptr(data); {
// Use device_ptr in Thrust sort thrust::device_ptr<double> dev_ptr(data);
// algorithm thrust::sort(dev_ptr, dev_ptr+N);
thrust::sort(dev_ptr, dev_ptr+N); }
} }

68
Calling Thrust from CUDA Fortran
Fortran interface to C wrapper using ISO C Bindings

module thrust
interface thrustsort
subroutine sort_int( input,N) & subroutine sort_float( input,N) &
bind(C,name="sort_int_wrapper") bind(C,name="sort_float_wrapper")
use iso_c_binding use iso_c_binding
integer(c_int),device:: input(*) real(c_float),device:: input(*)
integer(c_int),value:: N integer(c_int),value:: N
end subroutine end subroutine
end interface
subroutine sort_double( input,N) & end module
bind(C,name="sort_double_wrapper")
use iso_c_binding
real(c_double),device:: input(*)
integer(c_int),value:: N
end subroutine

69
CUDA Fortran Sorting with Thrust
program testsort !Print unsorted data
use thrust print *, cpuData
real, allocatable :: cpuData(:) !Send data to GPU
real, allocatable, device :: gpuData(:) gpuData = cpuData
integer :: N=10 !Sort the data
call thrustsort(gpuData,N)
!Allocate CPU and GPU arrays !Copy the result back
allocate(cpuData(N),gpuData(N)) cpuData = gpuData
!Fill the host array with random data !Print sorted data
do i=1,N print *, cpuData
cpuData(i)=random(i) !Deallocate arrays
end do deallocate(cpuData,gpuData)
end program testsort

nvcc -c -arch sm_20 csort.cu

pgf90 -Mcuda=cc20 -O3 -o testsort thrust_module.cuf testsort.cuf csort.o

$ ./testsort
Before sorting 4.1630346E-02 0.9124327 0.7832350 0.6540373 100.0000 0.3956419 0.2664442 0.13724658.0488138E-03 0.8788511
After sorting 8.0488138E-03 4.1630346E-02 0.1372465 0.26644420.3956419 0.6540373 0.7832350 0.87885110.9124327 100.0000

70
 CUDA Fortran Sorting with Thrust
program timesort ! Create events
use cudafor istat = cudaEventCreate ( startEvent )
use thrust istat = cudaEventCreate ( stopEvent )
real, allocatable :: cpuData(:) ! Send data to GPU
real, allocatable, device :: gpuData(:) gpuData = cpuData
integer:: N=100000000,istat !Sort the data
! cuda events for elapsing istat = cudaEventRecord ( startEvent , 0)
type (cudaEvent) :: startEvent, stopEvent call thrustsort(gpuData,N)
real :: time, random istat = cudaEventRecord ( stopEvent , 0)
istat = cudaEventSynchronize ( stopEvent )
!Allocate CPU and GPU arrays istat = cudaEventElapsedTime( time, &
allocate(cpuData(N), gpuData(N)) startEvent ,
!Fill the host array with random data stopEvent)
do i=1,N !Copy the result back
cpuData(i)=random(i) cpuData = gpuData
end do print *," Sorted array in:",time," (ms)"
print *,"After sorting", &
cpuData(1:5),cpuData(N-4:N)
!Deallocate arrays
deallocate(cpuData,gpuData)
$ ./timesort
Sorting array of 100000000 single precision end program timesort
Sorted array in: 194.6642 (ms)
After sorting 7.0585919E-09 1.0318221E-08 1.9398616E-08 3.1738640E-08 4.4078664E-08 0.9999999 0.9999999 1.000000 1.000000 1.000000

71
 Convolution Example

A B

Perform convolution of a stack of 2D arrays in frequency space

1. Send arrays to GPU memory
2. Transform arrays to frequency domain using 2D FFT from CUFFT
3. Perform point-wise complex multiplication and scaling (FFT(IFFT(A))=len(A)*A)
4. Transform result back to physical domain
5. Send result back to CPU memory

72
 Naive approach
Transfer A Transfer B FFT(A) FFT(B) conv IFFT(C) Transfer C

0.3s time

Smarter approach

I/O Transfer A Transfer B Transfer C

Compute FFT(A) FFT(B) conv IFFT(C)

time

Optimal approach
H2D A(1) B(1) A(2) B(2) A(N) B(N)
Compute FFT FFT conv IFFT FFT FFT conv IFFT FFT
A(N)
FFT
B(N)
conv IFFT
C(N)
A(1) B(1) C(1) A(2) B(2) C(2)

D2H C(1) C(2) C(N)

time
0.18s NOT TO SCALE

73
program driver do ifr=1,nomega
use cudafor ind = mod(ifr,num_streams)+1
use cufft
implicit none ! Send data to GPU
complex, allocatable,dimension(:,:,:), pinned :: A , B istat= cudaMemcpyAsync(A_d(1,1,ind),A(1,1,ifr),nxx*nyy, stream(ind))
complex, allocatable,dimension(:,:,:), device :: A_d, B_d istat= cudaMemcpyAsync(B_d(1,1,ind),B(1,1,ifr),nxx*nyy, stream(ind))

real:: elapsed_time,scale ! Execute FFTs on GPU

integer, parameter :: num_streams=8 call cufftSetStream(plan,stream(ind))
integer:: nxx, nyy, nomega, plan, stream(num_streams) call cufftExecC2C(plan ,A_d(1,1,ind),A_d(1,1,ind),CUFFT_FORWARD)
integer :: clock_start,clock_end,clock_rate, istat, ifr, i,j, ind call cufftExecC2C(plan ,B_d(1,1,ind),B_d(1,1,ind),CUFFT_FORWARD)
logical:: plog ! Convolution and scaling of the arrays
!$cuf kernel do(2) <<<*,*,stream=stream(ind)>>>
nxx=512; nyy=512 ; nomega=196 do j=1,nyy
do i=1,nxx
! Initialize FFT plan B_d(i,j,ind)= A_d(i,j,ind)*B_d(i,j,ind)*scale
call cufftPlan2d(plan,nyy,nxx,CUFFT_C2C) end do
end do
do i = 1,num_streams
istat= cudaStreamCreate(stream(i)) ! Execute FFTs on GPU
end do call cufftExecC2C(plan ,B_d(1,1,ind),B_d(1,1,ind),CUFFT_INVERSE)

! Copy results back

! Find the clock rate
istat=cudaMemcpyAsync( B(1,1,ifr),B_d(1,1,ind),nxx*nyy, stream=stream(ind))
call SYSTEM_CLOCK(COUNT_RATE=clock_rate)
end do
allocate(A(nxx,nyy,nomega),B(nxx,nyy,nomega)) istat=cudaThreadSynchronize()
allocate(A_d(nxx,nyy,num_streams),B_d(nxx,nyy,num_streams)) call SYSTEM_CLOCK(COUNT=clock_end) ! End timing
print *,"Elapsed time :", REAL(clock_end-clock_start)/REAL(clock_rate)
istat=cudaThreadSynchronize() deallocate(A,B,A_d,B_d)
call SYSTEM_CLOCK(COUNT=clock_start) ! Start timing
scale =1./(nxx*nyy) end program

74
Computing π with CUDA Fortran

π = 4 ∗ (Σ red points)/(Σ points)

Simple example:
–Generate random numbers ( CURAND)
–Compute sum using of kernel loop directive
–Compute sum using two stages reduction with Cuda Fortran kernels
–Compute sum using single stage reduction with Cuda Fortran kernel
–Accuracy

75
CUDA Libraries from CUDA Fortran
• All the toolkit libraries have C interfaces
• Use F90 interfaces and ISO C Binding to use from CUDA Fortran
interface curandGenerateUniform
!curandGenerateUniform(curandGenerator_t generator, float *outputPtr, size_t num);
subroutine curandGenerateUniform(generator, odata, numele) &
bind(C,name='curandGenerateUniform')
use iso_c_binding
integer(c_size_t),value:: generator
!pgi$ ignore_tr odata
real(c_float), device:: odata(*)
integer(c_size_t),value:: numele
end subroutine curandGenerateUniform

!curandGenerateUniformDouble(curandGenerator_t generator, double *outputPtr, size_t num);

subroutine curandGenerateUniformDouble(generator, odata, numele) &
bind(C,name='curandGenerateUniformDouble')
use iso_c_binding
integer(c_size_t),value:: generator
!pgi$ ignore_tr odata
real(c_double), device:: odata(*)
integer(c_size_t),value:: numele
end subroutine curandGenerateUniformDouble
end interface curandGenerateUniform

76
 Computing π with CUF Kernel
! Perform the test on GPU using CUF kernel
! Compute pi using a Monte Carlo method
inside=0
program compute_pi
!$cuf kernel do <<<*,*>>>
use precision
do i=1,Nhalf
use cudafor ! CUDA Fortran runtime
if( (deviceData(i)**2+deviceData(i+Nhalf)**2) &
use curand ! CURAND interface
<= 1._fp_kind ) inside=inside+1
implicit none
end do
real(fp_kind), allocatable, pinned:: hostData(:)
! Perform the test on CPU
real(fp_kind), allocatable, device:: deviceData(:)
inside_cpu=0
real(fp_kind):: pival
do i=1,Nhalf
integer :: inside_cpu,inside, i, iter, Nhalf
if( (hostData(i)**2+hostData(i+Nhalf)**2) &
integer(kind=8) :: gen, N, seed=1234
<= 1._fp_kind) inside_cpu=inside_cpu+1
end do
! Define how many numbers we want to generate
! Check the results
N=2000
if (inside_cpu .ne. inside) &
Nhalf=N/2
print *,"Mismatch between CPU/ &
! Allocate arrays on CPU and GPU
GPU",inside_cpu,inside
allocate(hostData(N), deviceData(N))
! Print the value of pi and the error
! Create pseudonumber generator
pival= 4._fp_kind*real(inside,fp_kind) &
call curandCreateGenerator(gen, &
/real(Nhalf,fp_kind)
CURAND_RNG_PSEUDO_DEFAULT)
print"(t3,a,i10,a,f10.8,a,e11.4)", "Samples=", &
! Set seed
Nhalf," Pi=", pival," Error=", &
call curandSetPseudoRandomGeneratorSeed( gen, seed)
abs(pival-2.0_fp_kind*asin(1.0_fp_kind))
! Generate N floats or double on device
! Deallocate data on CPU and GPU
call curandGenerateUniform(gen, deviceData, N)
deallocate(hostData,deviceData)
! Copy the data back to CPU to check result later
! Destroy the generator
hostData=deviceData
call curandDestroyGenerator(gen)
end program compute_pi
77
 Computing π
pgf90 -O3 -Mpreprocess -o pi_gpu precision_module.cuf curand_module.cuf pi.cuf -lcurand

Compute pi in single precision (seed 1234)

Samples= 10000 Pi=3.11120009 Error= 0.3039E-01
Samples= 100000 Pi=3.13632011 Error= 0.5273E-02
Samples= 1000000 Pi=3.14056396 Error= 0.1029E-02
Samples= 10000000 Pi=3.14092445 Error= 0.6683E-03
Samples= 100000000 Pi=3.14158082 Error= 0.1192E-04

78
 Computing π
pgf90 -O3 -Mpreprocess -o pi_gpu precision_module.cuf curand_module.cuf pi.cuf -lcurand

Compute pi in single precision (seed 1234)

Samples= 10000 Pi=3.11120009 Error= 0.3039E-01
Samples= 100000 Pi=3.13632011 Error= 0.5273E-02
Samples= 1000000 Pi=3.14056396 Error= 0.1029E-02
Samples= 10000000 Pi=3.14092445 Error= 0.6683E-03
Samples= 100000000 Pi=3.14158082 Error= 0.1192E-04

Where is the difference coming from?

if( ( hostData(i)**2+ hostData(i+Nhalf)**2) <= 1._fp_kind) inside_cpu=inside_cpu+1 (CPU)
if( (deviceData(i)**2+deviceData(i+Nhalf)**2) <= 1._fp_kind ) inside=inside+1 (GPU)

- Sum of the point inside the circle is done with integers ( no issues due to floating point arithmetic)
- Computation of the distance from the origin (x*x+y*y), no special functions just + and *

78
 Computing π
pgf90 -O3 -Mpreprocess -o pi_gpu precision_module.cuf curand_module.cuf pi.cuf -lcurand

Compute pi in single precision (seed 1234) Compute pi in single precision (seed 1234567)
Samples= 10000 Pi=3.11120009 Error= 0.3039E-01 Samples= 10000 Pi=3.16720009 Error= 0.2561E-01
Samples= 100000 Pi=3.13632011 Error= 0.5273E-02 Samples= 100000 Pi=3.13919997 Error= 0.2393E-02
Samples= 1000000 Pi=3.14056396 Error= 0.1029E-02 Samples= 1000000 Pi=3.14109206 Error= 0.5007E-03
Samples= 10000000 Pi=3.14092445 Error= 0.6683E-03 Samples= 10000000 Pi=3.14106607 Error= 0.5267E-03
Samples= 100000000 Pi=3.14158082 Error= 0.1192E-04 Mismatch between CPU/GPU 78534862 78534859
Samples= 100000000 Pi=3.14139414 Error= 0.1986E-03

Where is the difference coming from?

if( ( hostData(i)**2+ hostData(i+Nhalf)**2) <= 1._fp_kind) inside_cpu=inside_cpu+1 (CPU)
if( (deviceData(i)**2+deviceData(i+Nhalf)**2) <= 1._fp_kind ) inside=inside+1 (GPU)

- Sum of the point inside the circle is done with integers ( no issues due to floating point arithmetic)
- Computation of the distance from the origin (x*x+y*y), no special functions just + and *

78
GPU Accuracy
• FERMI GPUs are IEEE-754 compliant, both for single and double precision
• Support for Fused Multiply-Add instruction (IEEE 754-2008)
• Results with FMA could be different* from results without FMA
• In CUDA Fortran is possible to toggle FMA on/off with a compiler switch:

-Mcuda=nofma
• Extremely useful to compare results to “golden” CPU output
• FMA is being supported in future CPUs
Compute pi in single precision (seed=1234567 FMA disabled)
Samples= 10000 Pi=3.16720009 Error= 0.2561E-01
Samples= 100000 Pi=3.13919997 Error= 0.2393E-02
Samples= 1000000 Pi=3.14109206 Error= 0.5007E-03
Samples= 10000000 Pi=3.14106607 Error= 0.5267E-03
Samples= 100000000 Pi=3.14139462 Error= 0.1981E-03

*GPU results with FMA are identical to CPU if operations are done in double precision

79
Reductions on GPU 3 1 7 0 4 1 6 3

4 7 5 9
• Parallelism across blocks 11 14
• Parallelism within a block 25
• No global synchronization
– two-stage approach (two kernel lauches), same code for both stages

3 1 7 0 4 1 6 3 3 1 7 0 4 1 6 3 3 1 7 0 4 1 6 3 3 1 7 0 4 1 6 3 3 1 7 0 4 1 6 3 3 1 7 0 4 1 6 3 3 1 7 0 4 1 6 3 3 1 7 0 4 1 6 3
4 7 5 9 4 7 5 9 4 7 5 9 4 7 5 9 4 7 5 9 4 7 5 9 4 7 5 9 4 7 5 9
11 14 11 14 11 14 11 14 11 14 11 14 11 14 11 14
25 25 25 25 25 25 25 25 Level 0:
8 blocks

3 1 7 0 4 1 6 3 Level 1:
4 7 5 9
11
25
14 1 block

80
Parallel Reduction: Sequential Addressing
10 1 8 -1 0 -2 3 5 -2 -3 2 7 0 11 0 2

Step 1 Thread
Stride 8 IDs
1 2 3 4 5 6 7 8
Values 8 -2 10 6 0 9 3 7 -2 -3 2 7 0 11 0 2

Step 2 Thread
Stride 4 IDs 1 2 3 4
Values 8 7 13 13 0 9 3 7 -2 -3 2 7 0 11 0 2

Step 3 Thread
Stride 2 IDs 1 2
Values 21 20 13 13 0 9 3 7 -2 -3 2 7 0 11 0 2

Step 4 Thread
Stride 1 IDs 1

Values 41 20 13 13 0 9 3 7 -2 -3 2 7 0 11 0 2

81
Computing π with CUDA Fortran Kernels
attributes(global) subroutine partial_sum psum(index)=interior
(input,partial,N)
call syncthreads()
real(fp_kind) :: input(N)
integer :: partial(256)
inext=blockDim%x/2
integer, shared, dimension(256) :: psum
do while ( inext >=1 )
integer(kind=8),value :: N
if (index <=inext) psum(index) = &
integer :: i,index, inext,interior
psum(index) + psum(index+inext)
index=threadIdx%x+(BlockIdx%x-1)*BlockDim%x inext = inext /2
! Check if the point is inside the circle call syncthreads()
! and increment local counter end do
interior=0
do i=index,N/2,BlockDim%x*GridDim%x ! Each block writes back its partial sum
if( (input(i)**2+input(i+N/2)**2) <= 1._fp_kind) & if (index == 1) partial(BlockIdx%x)=psum(1)
interior=interior+1 end subroutine
end do

! Local reduction per block ! Compute the partial sums with 256 blocks of 512 threads
index=threadIdx%x
call partial_sum<<<256,512,512*4>>>(deviceData,partial,N)
! Compute the final sum with 1 block of 256 threads
call final_sum<<<1,256,256*4>>>(partial,inside_gpu)

82
Computing π with CUDA Fortran Kernels
attributes(global) subroutine final_sum(partial,nthreads,total)
integer, intent(in) :: partial(nthreads)
integer, intent(out) :: total
integer, shared :: psum(*)
integer :: index, inext

index=threadIdx%x

! load partial sums in shared memory

psum(index)=partial(index)
call syncthreads()

inext=blockDim%x/2
do while ( inext >=1 )
if (index <=inext) psum(index)=psum(index)+psum(index+inext)
inext = inext /2
call syncthreads()
end do

! First thread has the total sum, writes it back to global memory
if (index == 1) total=psum(1)
end subroutine

83
 Computing π with an Atomic Lock
Instead of storing back the partial sum:
! Each block writes back its partial sum
if (index == 1) partial(BlockIdx%x)=psum(1)

use an atomic lock to ensure that one block at the time updates the final sum:

if (index == 1) then
do while ( atomiccas(lock,0,1) == 1) !set lock
end do
partial(1)=partial(1)+psum(1) ! atomic update of partial(1)
call threadfence() ! Wait for memory transaction to be visible to all the other threads
lock =0 ! release lock
end if

partial(1)=0
call sum<<<64,256,256*4>>>(deviceData,partial,N)
inside=partial(1)

84