Lecture 9(B): GPUs & GPGPU

Transcription

1 Lecture 9(B): GPUs & GPGPU John-Philip Taylor 26 March 2015 Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)

2 Outline OpenGL Primitives and Vertices Vertex Shader Rasteriser Fragment Shader OpenCL Conceptually Initialisation Matrix Multiplication Further Reading Disclaimers and Copyright

3 OpenGL

4 Primitives and Vertices There are 3 types of primitives: Points Lines Triangles They are described by vertices

5 Primitives and Vertices There are 3 types of primitives: Points Lines Triangles They are described by vertices

6 Primitives and Vertices There are 3 types of primitives: Points Lines Triangles They are described by vertices

7 Vertex Attribute Tables X Y S T R G B These attribute data is stored in a vertex buffer, which is a one-dimensional structure. The order of the attributes within this buffer is user-defined. Tell the GPU where to find what by calling glvertexattribpointer, which uses start and stride indexing.

8 Vertex Shader #version 330 in vec2 vposition; // X, Y from vertex buffer in vec2 vtexture; // S, T from vertex buffer in vec3 vcolour; // R, G, B from vertex buffer out vec2 ftexture; // To fragment shader out vec3 fcolour; // To fragment shader void main(void){ // Main function ftexture = vtexture; fcolour = vcolour; gl_position = vec4(vposition.xy, 0.0, 1.0); }

9 Vertex Shader Vertex 1 Vertex 2 Vertex 3 Vertex 4 In lockstep Shader Instance Shader Instance Shader Instance Shader Instance In order Triangle 1 Rasteriser Triangle 2 Rasteriser

10 Rasteriser Unit Cube (1, 1, -1) Screen Space (1, 1) (-1, -1, 1) (-1, -1) To Fragment Shader

11 Rasteriser Unit Cube (1, 1, -1) Screen Space (1, 1) (-1, -1, 1) (-1, -1) To Fragment Shader All vertex shader output variables are linearly interpolated over the primitive area.

12 Fragment Shader #version 330 in vec2 ftexture; in vec3 fcolour; // From vertex shader uniform sampler2d Texture; // The texture handle out vec4 FragColour; // To frame buffer void main(void){ FragColour = 0.5*texture2D(Texture, ftexture) + 0.5*vec4 (fcolour, 1.0 ); } This is run, in lockstep, for every fragment in the primitive. The output colour is depth-tested and blended with the colour currently in the frame buffer.

13 OpenCL

14 Conceptually Input Buffer Input Buffer Input Buffer Kernel Instances Output Buffer Output Buffer Output Buffer

15 Conceptually Similar to vertex buffers and textures Input Buffer Input Buffer Input Buffer Similar to vertex and fragment shaders Kernel Instances Similar to depth and frame buffers Output Buffer Output Buffer Output Buffer

16 Memory Model Work Group Worker Worker Private Private Local Work Group Worker Worker Private Private Local Global

17 Initialisation 1. Get platform ID Many platforms exist on any one PC. Choose between CPU, GPU, FPGA and any other OpenCL platform the system supports. 2. Compile kernels 3. Create buffers 4. Create command queue 5. Enqueue write 6. Enqueue kernel run 7. Enqueue read 8. Wait for queue to finish

18 Initialisation 1. Get platform ID 2. Compile kernels Similar to OpenGL shaders, the OpenCL driver compiles the kernels. 3. Create buffers 4. Create command queue 5. Enqueue write 6. Enqueue kernel run 7. Enqueue read 8. Wait for queue to finish

19 Initialisation 1. Get platform ID 2. Compile kernels 3. Create buffers Buffers reside on the target system, not necessarily in CPU RAM. 4. Create command queue 5. Enqueue write 6. Enqueue kernel run 7. Enqueue read 8. Wait for queue to finish

20 Initialisation 1. Get platform ID 2. Compile kernels 3. Create buffers 4. Create command queue The CPU enqueues commands to be executed in FIFO order. The CPU can then carry on with other work while the OpenCL system is busy. 5. Enqueue write 6. Enqueue kernel run 7. Enqueue read 8. Wait for queue to finish

21 Initialisation 1. Get platform ID 2. Compile kernels 3. Create buffers 4. Create command queue 5. Enqueue write Copy the data from CPU RAM to the OpenCL system global memory. 6. Enqueue kernel run 7. Enqueue read 8. Wait for queue to finish

22 Initialisation 1. Get platform ID 2. Compile kernels 3. Create buffers 4. Create command queue 5. Enqueue write 6. Enqueue kernel run Specify global size (everything) Specify local size (work group size) Global size must be an integer multiple of local size 7. Enqueue read 8. Wait for queue to finish

23 Initialisation 1. Get platform ID 2. Compile kernels 3. Create buffers 4. Create command queue 5. Enqueue write 6. Enqueue kernel run 7. Enqueue read Copy the data from the OpenCL system global memory to CPU RAM. 8. Wait for queue to finish

24 Initialisation 1. Get platform ID 2. Compile kernels 3. Create buffers 4. Create command queue 5. Enqueue write 6. Enqueue kernel run 7. Enqueue read 8. Wait for queue to finish Wait for the OpenCL system to finish performing the commands in the queue.

25 Matrix Multiplication Example A B Y

26 C Implementation void Multiply( float** A, // Input matrix float** B, // Input matrix float** Y, // Output matrix int N // Matrix size ){ int i, j, k; for(i = 0; i < N; i++){ for(j = 0; j < N; j++){ Y[i][j] = 0; for(k = 0; k < N; k++){ Y[i][j] += A[i][k] * B[k][j]; }}}}

27 OpenCL Implementation kernel void Multiply( global float* A, // Global input buffer global float* B, // Global input buffer global float* Y, // Global output buffer const int N // Global uniform ){ const int i = get_global_id(0); // 1st dimension index const int j = get_global_id(1); // 2nd dimension index // Private variables int k; float f = 0.0; for(k = 0; k < N; k++) f += A[i*N + k] * B[k*N + j]; Y[i*N + j] = f; }

28 Further Reading R Wright, N Haemel and G Sellers OpenGL SuperBible, 6 th ed Addison-Wesley, 2014, ISBN A Munshi, B R Gaster, T G Mattson, J Fung and D Ginsburg OpenCL Programming Guide Addison-Wesley, 2012, ISBN Mac Developer Library OpenCL Hello World Example Altera OpenCL SDK for FPGA

29 Disclaimers and Copyright I have tried to follow the correct practices concerning copyright and licensing of material, particularly image sources that have been used in this presentation. I have put much effort into trying to make this material open access so that it can be of benefit to others in their teaching and learning practice. Any mistakes or omissions with regards to these issues I will correct when notified. To the best of my understanding the material in these slides can be shared according to the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license, and that is why I selected that license to apply to this presentation (it s not because I particularly want my slides referenced but more to acknowledge the sources and generosity of others who have provided free material such as the images I have used). Image sources: Wikipedia (open commons)