Introduction to CUDA C

More documents

Recommendations

Info

Sharing Data Between Threads � Terminology: A block of threads shares memory called…shared memory � Extremely fast, on-chip memory (user-managed cache) � Declared with the __shared__ <strong>CUDA</strong> keyword � Not visible <strong>to</strong> threads in other blocks running in parallel Block 0 Threads Shared Memory Block 1 Threads Shared Memory Block 2 Threads Shared Memory …
Parallel Dot Product: dot() � We perform parallel multiplication, serial addition: #define N 512 __global__ void dot( int *a, int *b, int *c ) { // Shared memory for results of multiplication __shared__ int temp[N]; temp[threadIdx.x] = a[threadIdx.x] * b[threadIdx.x]; } // Thread 0 sums the pairwise products if( 0 == threadIdx.x ) { int sum = 0; for( int i = 0; i < N; i++ ) sum += temp[i]; *c = sum; }
Page 1 and 2: Introduction to CUDA C San Jose Con
Page 3 and 4: What is CUDA? � CUDA Architecture
Page 5 and 6: CUDA C Prerequisites � You (proba
Page 7 and 8: Hello, World! int main( void ) { pr
Page 9 and 10: Hello, World! with Device Code __gl
Page 11 and 12: A More Complex Example � A simple
Page 13 and 14: Memory Management � Host and devi
Page 15 and 16: A More Complex Example: main() int
Page 17 and 18: Parallel Programming in CUDA C �
Page 19 and 20: Parallel Programming in CUDA C �
Page 21 and 22: Parallel Addition: main() #define N
Page 23 and 24: Review � Difference between “ho
Page 25 and 26: Threads � Terminology: A block ca
Page 27 and 28: Parallel Addition (Threads): main()
Page 29 and 30: Indexing Arrays With Threads And Bl
Page 31 and 32: Addition with Threads and Blocks
Page 33 and 34: Parallel Addition (Blocks/Threads):
Page 35 and 36: Dot Product � Unlike vector addit
Page 37: Dot Product � But we need to shar
Page 41 and 42: Faulty Dot Product Exposed! � Ste
Page 43 and 44: Synchronization � We need threads
Page 45 and 46: Parallel Dot Product: dot() __globa
Page 47 and 48: Parallel Dot Product: main() } // c
Page 49 and 50: Review (cont) � Using __shared__
Page 51 and 52: Multiblock Dot Product: Algorithm
Page 53 and 54: Multiblock Dot Product: dot() #defi
Page 55 and 56: Global Memory Contention Block 0 su
Page 57 and 58: Atomic Operations � Terminology:
Page 59 and 60: Parallel Dot Product: main() #defin
Page 61 and 62: Review � Race conditions — Beha
Page 63: Questions � First my questions

Introduction to CUDA C

Create successful ePaper yourself

Delete template?

Save as template?