Ejercicios Planteados

Anuncio
Programación Gráfica de Altas Prestaciones
Ejercicio práctico 15/04/2015
Documents at your the corresponding account in hgpu.ugr.es
You will find the following folder in your home folder:
advection1d_code
and the following files in this folder: advection1d.cu and Makefile.
File advection1d.cu contains C+ CUDA C code of a linear advection solver. It
includes the following:
1) The declaration of two kernels to perform the FD update (FD_kernel1 and
FD_kernel2).
2) A GPU version of the solver (based on FD_kernel1).
3) A CPU version of the solver
4) Code to check the performance and accuracy.
File Makefile is used to automatically compile and link the source code to
generate the executable file (advection1d).
Generating the executable file and executing the solver
The Makefile will be activated by executing the make command in the terminal
which generates a call to nvcc with the suitable choices and argumnents:
> make
nvcc -O3 -m64 -gencode arch=compute_35,code=sm_35 advection1d.cu -o advection1d
Once we have built the executable file (advection1d), we can execute it by
writing:
./advection1d
ASSIGNMENTS FOR THE PARTICIPANTS
These assignments involves to modify the file advection1d.cu. Therefore, it is
recommendable to copy the filename with a different name to avoid the loss of
information. So for example, we could copy the original file using another name
with:
> cp advectiontion1d.cu advectiontion1d_orig.cuand later perform
modifications on advectiontion1d.cu.
The work performed
(1) Run the linear advection CUDA solver by varying the number of mesh
points (the actual number of mesh points is n+1 where n is an integer
variable) and the BLOCKSIZE (currently is 256 but you could test with 64,
128, 512, 1024,...). The goal is to study how these parameters affects the
solver performance and to obtain conclusions.
(2) In the linear advection CUDA solver, each thread just computes one value
of the new vector state d_phi_new (see FD_kernel1 definition). Modify the
definition of the kernel FD_kernel1 (global memory version of the FD
Update) and the kernel launch in the file advection1d.cu, in order to
impose that each CUDA thread computes two contiguous point values for
the new state vector d_phi_new.
(3) Modify the advectiontion1d.cu file to use the tiled version of the linear
advection CUDA solver. The definition of the tiled kernel is included in the
file ( kernel FD_kernel2). Study how is the solver performance affected by
the new kernel and give your conclusions.
The conclusions obtained from points (1) and (3) and the kernel Description
obtained in (2) must be sent to [email protected].
Guía rápida de emacs
Alterando el texto
Ctrl k – kill, “mata” el texto. Esto implica que lo borra, pero lo mantiene guardado y puede ser
recuperado con Ctrl Y.
Ctrl y – “Yanks text”, pega el texto matado o cortado/copiado con w.
Guardar y salir
Ctrl x Ctrl s – Guardar un archivo
Ctrl X Ctrl C – Salir de Emacs (pregunta si no has guardado)
Ctrl S – Buscar strings incrementalmente (con Ctrl S sigo buscando hacia adelante, con Ctrl R
busco hacia atrás y con enter termino la busqueda)
Mover por páginas
Ctrl v – (siguiente)
Alt v – (anterior)
Ctrl l – Limpiar la pantalla y volver a mostrar todo el texto, moviendo el texto alrededor del cursor
al centro de la pantalla.
Moverse entre el texto
Ctrl n – Moverse a la siguiente línea
Ctrl p – Moverse a la línea anterior
Ctrl a – Moverse al comienzo de una línea
Ctrl e – Moverse al final de una línea
Alt a – Moverse al inicio de una oración
Alt e – Moverse al final de una oración
Descargar