Why should I bother GPU?

Moore is no more

Real parallel computation

Available for normal people

Applications

• Graphics
• Simulations
• AI
• Data Analysis

“Talk is cheep show me the code”

Linus Torvalds

Add the corresponding locations of A and B, and store the result in C.

Plain Old C

							void vecadd( int *A , int *B , int *C)
							{
							    for (int i = 0; i < L; i++) {
							        C[i] = A[i] + B[i];
							    }
							}
						

OpenMP

							void vecadd( int *A , int *B , int *C)
							{
							    chunk = CHUNKSIZE;
							    #pragma omp parallel shared(A,B,C,chunk) private(i)
							    {
							        #pragma omp for schedule(dynamic,chunk) nowait
							        for (int i = 0; i < L; i++) {
							            C[i] = A[i] + B[i];
							        }
							    }
							}
							

GLSL

							#version 110

							uniform sampler2D texture1;
							uniform sampler2D texture2;

							void main() {
							    vec4 A = texture2D(texture1, gl_TexCoord[0].st);
							    vec4 B = texture2D(texture2, gl_TexCoord[0].st);
							    gl_FragColor = A + B;
							}
							

OpenCL

								__kernel
								void vecadd(__global int *A,
								            __global int *B,
								            __global int *C)
								{
								    int id = get_global_id(0);
								    C[id] = A[id] + B[id];
								}
							

CUDA

								__global__
								void vecadd( int *A , int *B , int *C)
								{
								    int id = blockIdx.x*blockDim.x+threadIdx.x;
								    C[id] = A[id] + B[id] ;
								}
							

Not so fast

1. Copy data to GPU
2. Launch kernel
3. Check for errors
4. Copy output back to RAM

Results

http://hpclab.blogspot.com/2011/09/is-gpu-good-for-large-vector-addition.html

but for harder problems

http://hpclab.blogspot.com/2011/09/is-gpu-good-for-large-vector-addition.html

How it works inside?

Rules Of Thumb

RTFM

“Life is too short for man pages, and occasionally much too short without them.”

Randall Munroe (xkcd.com)

Think parallel

SIMD

Problems

Problem: Development is hard

Solution: Always have spare GPU in your computer

Problem: Debugging is impossible

Solution: Write tests and run them!

Problem: Copying data to/from GPU is slow

Solution: Use stream and compute while data are loaded

Problem: GPU doesn't like 64bit computation

Solution: Wait for next release

Problem: I dont want to code a lot

Solution: Use libs

• ArrayFire
• Thrust (STL for CUDA)
• cuBLAS (Basic Linear Algebra Subprograms)
• cuFFT
• cuDNN (GPU-accelerated library of primitives for deep neural networks)

Before you code your custom solution.

PGStorm

							postgres=# SELECT COUNT(*) FROM t1 WHERE sqrt((x-25.6)^2 + (y-12.8)^2) < 15;
							count
							-------
							6718
							(1 row)

							Time: 7019.855 ms
						

							postgres=# SELECT COUNT(*) FROM t2 WHERE sqrt((x-25.6)^2 + (y-12.8)^2) < 15;
							count
							-------
							6718
							(1 row)

							Time: 176.301 ms
						

t1 and t2 contain same contents with 10 millions of records, but t1 is a regular table and t2 is a foreign table managed by PG-Strom