1. GPU Programming
Roberto Bonvallet
Departamento de Inform´ tica
a
Universidad T´ cnica Federico Santa Mar´a
e ı
Junio de 2010

2. CPU vs GPU peak performance

3. CPU and GPU architectures
Control ALU ALU
ALU ALU
Cache
DRAM DRAM

4. CPU and GPU architectures
DRAM

5. CPU and GPU architectures

6. Nvidia Tesla architecture

7. Task and data parallelism

8. Task and data parallelism
Task parallelism:
distributed
processing
distributed memory
message passing

9. Task and data parallelism
Task parallelism:
distributed
processing
distributed memory
message passing
Data parallelism:
same instruction on
different data
shared memory

10. Thread and memory hierarchies
Thread hierarchy:

11. Thread and memory hierarchies
Thread hierarchy:
grid of blocks

12. Thread and memory hierarchies
Thread hierarchy:
grid of blocks
blocks of threads

13. Thread and memory hierarchies
Thread hierarchy:
grid of blocks
blocks of threads
Memory hierarchy:
global memory (large, slow)
shared memory (per-block, small, fast)
registers (per-thread, small, fast)

14. Matrix-matrix multiplication

15. Matrix-matrix multiplication
cij = aik bkj
k

16. Matrix-matrix multiplication
cij = aik bkj
k
Cij = Aik Bkj
k

17. Matrix-matrix multiplication
cij = aik bkj
k
Cij = Aik Bkj
k
Multiplication kernel:
initialize element of
Cij = 0
for each k:
fetch element of
Aik , Bkj into shared
memory
synchronize
compute element
of Cij = Cij + Aik Bkj
synchronize

18. Nvidia C1060
Core clock 602 Mhz
Multiprocessors 30
Thread processors 240 = 30 × 8
Memory size 4 GB
Memory bandwidth 102.4 GB/s
Single precision pp 933.12 Gflop
Double precision pp 77.76 Gflop

19. CUDA programming
Array allocation and copying
cudaMalloc((void **) &p, mem_size);
cudaMemcpy(host_p, dev_p, mem_size,
cudaMemcpyHostToDevice);
[...]
cudaMemcpy(dev_p, host_p, mem_size,
cudaMemcpyDeviceToHost);
cudaFree(p);

20. CUDA programming
Kernel definition
__global__ void
vector_sum(float *a, float *b, float *c) {
int i = blockIdx.x * blockDim.x +
threadIdx.x;
c[i] = a[i] + b[i];
}

21. CUDA programming
Kernel definition
__global__ void
vector_sum(float *a, float *b, float *c) {
int i = blockIdx.x * blockDim.x +
threadIdx.x;
c[i] = a[i] + b[i];
}
Kernel launch
f<<<grid_size, block_size,
sh_mem_size>>>(a, b, c);

22. Vortex Methods
Fluid discretized as vortices
(x, y, α)

23. Vortex Methods
Fluid discretized as vortices
(x, y, α)
Vortex interaction:
1
K(x, y) = − (−y, x)
2π x

24. Vortex Methods
Fluid discretized as vortices
(x, y, α)
Vortex interaction:
1
K(x, y) = − (−y, x)
2π x
Biot-Savart law:
u(x) = αp K(x − xp )
p

25. GPU velocity evaluation