GPU Programming Introduction

Transcription

1 GPU Programming Introduction DR. CHRISTOPH ANGERER, NVIDIA

2 AGENDA Introduction to Heterogeneous Computing Using Accelerated Libraries GPU Programming Languages Introduction to CUDA Lunch

3 What is Heterogeneous Computing? Application Execution CPU High Serial Performance High Data Parallelism GPU +

4 Low Latency or High Throughput?

5 Latency vs. Throughput F-22 Raptor 1500 mph Knoxville to San Jose in 1:25 Seats 1 Boeing mph Knoxville to San Jose in 4:20 Seats 200

6 Latency vs. Throughput F-22 Raptor Latency 1:25 Throughput 1 / 1.42 hours = 0.7 people/hr. Boeing 737 Latency 4:20 Throughput 200 / 4.33 hours = 46.2 people/hr.

7 Low Latency or High Throughput? CPU architecture must minimize latency within each thread GPU architecture hides latency with computation from other threads CPU core Low Latency Processor Computation Thread/Warp T 1 T 2 T 3 T 4 GPU Streaming Multiprocessor High Throughput Processor W 4 W 3 W 2 W 1 T n Processing Waiting for data Ready to be processed Context switch

8 Anatomy of a GPU-Accelerated Application Serial code executes in a Host (CPU) thread Parallel code executes in many Device (GPU) threads across multiple processing elements Application Serial code Host = CPU Parallel code Device = GPU Serial code Host = CPU Parallel code Device = GPU...

9 Simple Processing Flow PCI Bus 1. Copy input data from CPU memory to GPU memory

10 Simple Processing Flow PCI Bus 1. Copy input data from CPU memory to GPU memory 2. Load GPU code and execute it

11 Simple Processing Flow PCI Bus 1. Copy input data from CPU memory to GPU memory 2. Load GPU code and execute it 3. Copy results from GPU memory to CPU memory

12 3 APPROACHES TO GPU PROGRAMMING Applications Libraries Compiler Directives Programming Languages Easy to use Most Performance Easy to use Portable code Most Performance Most Flexibility

13 Simplicity SIMPLICITY & PERFORMANCE Accelerated Libraries Little or no code change for standard libraries, high performance. Limited by what libraries are available Compiler Directives Based on existing programming languages, so they are simple and familiar. Performance may not be optimal because directives often do not expose low level architectural details Performance Parallel Programming languages Expose low-level details for maximum performance Often more difficult to learn and more time consuming to implement.

14 Libraries

15 Libraries: Easy, High-Quality Acceleration Ease of use: Drop-in : Quality: Performance: Using libraries enables GPU acceleration without in-depth knowledge of GPU programming Many GPU-accelerated libraries follow standard APIs, thus enabling acceleration with minimal code changes Libraries offer high-quality implementations of functions encountered in a broad range of applications NVIDIA libraries are tuned by experts

16 Gpu accelerated libraries Providing Drop-in Acceleration Linear Algebra FFT, BLAS, SPARSE, Matrix NVIDIA cufft, cublas, cusparse Numerical & Math RAND, Statistics NVIDIA Math Lib NVIDIA curand Data Struct. & AI Sort, Scan, Zero Sum Visual Processing Image & Video NVIDIA NPP GPU AI Board Games NVIDIA Video Encode GPU AI Path Finding

17 3 Steps to CUDA-accelerated application Step 1: Substitute library calls with equivalent CUDA library calls saxpy ( ) cublassaxpy ( ) Step 2: Manage data locality - with CUDA: cudamalloc(), cudamemcpy(), etc. - with CUBLAS: cublasalloc(), cublassetvector(), etc. Step 3: Rebuild and link the CUDA-accelerated library nvcc myobj.o l cublas

18 Programming Languages (CUDA)

19 GPU Programming Languages Numerical analytics Fortran C C++ Python C# MATLAB, Mathematica, LabVIEW CUDA Fortran CUDA C Thrust, CUDA C++ PyCUDA, Copperhead GPU.NET

20 What is CUDA? CUDA Platform and Programming Model Expose GPU computing for general purpose A model how to offload work to the GPU and how the work is executed on the GPU CUDA C/C++ Based on industry-standard C/C++ Small set of extensions to enable heterogeneous programming Straightforward APIs to manage devices, memory etc.

21 Anatomy of a CUDA Application Serial code executes in a Host (CPU) thread (Functions) Parallel code executes in many Device (GPU) threads across multiple processing elements (Kernels) CUDA Application Serial code Host = CPU Parallel code Device = GPU Serial code Host = CPU Parallel code Device = GPU...

22 CUDA Kernels: Parallel Threads A kernel is a function executed on the GPU as an array of threads in parallel All threads execute the same code can take different paths but the fewer divergence between neighboring threads the better global void mykernel( float * input, float * output) { float x = input[threadidx.x]; float y = func(x); output[threadidx.x] = y; } Each thread has an ID Select input/output data Control decisions

23 CUDA Kernels: Blocks Threads are grouped into blocks Blocks are grouped into a grid

24 CUDA Kernels: Grids blockidx.z blockidx.y 2 GPU 1 A grid is a 1D, 2D or 3D index space Threads query their position inside the index space using blockidx.{x,y,z} and threadidx.{x,y,z} and decide on the work they have to do Only threads in the same block can communicate directly (via shared memory) blockidx.x

25 HELLO WORLD!

26 Hello World! int main(void) { } printf("hello World!\n"); return 0; Standard C that runs on the host NVIDIA compiler (nvcc) can be used to compile programs with no device code Output: $ module load cudatoolkit $ nvcc hello_world.cu $ a.out Hello World!

27 Hello World! with Device Code global void mykernel(void) { } int main(void) { mykernel<<<1,1>>>(); printf("hello World!\n"); return 0; } Two new syntactic elements

28 Hello World! with Device Code global void mykernel(void) { } CUDA C/C++ keyword global indicates a function that: Runs on the device Is called from host code nvcc separates source code into host and device components Device functions (e.g. mykernel()) processed by NVIDIA compiler Host functions (e.g. main()) processed by standard host compiler - gcc, cl.exe

29 Hello World! with Device Code mykernel<<<1,1>>>(); Triple angle brackets mark a call from host code to device code Also called a kernel launch We ll return to the parameters (1,1) in a moment That s all that is required to execute a function on the GPU!

30 Parallel Programming in CUDA C/C++ But wait GPU computing is about massive parallelism! We need a more interesting example We ll start by adding two integers and build up to vector addition a b c

31 Addition on the Device A simple kernel to add two integers global void add(int a, int b, int *c) { } *c = a + b; As before global is a CUDA C/C++ keyword meaning add() will execute on the device add() will be called from the host Note that we are passing pointers. - Necessary for c because we want the result back

32 Memory Management Host and device memory are separate entities Device pointers point to GPU memory May be passed to/from host code May not be dereferenced in host code Host pointers point to CPU memory May be passed to/from device code May not be dereferenced in device code Simple CUDA API for handling device memory cudamalloc(), cudafree(), cudamemcpy() Similar to the C equivalents malloc(), free(), memcpy() NEW (CUDA 6.0): Unified Memory can do the memory management for you!

33 Addition on the Device: add() Returning to our add() kernel global void add(int a, int b, int *c) { } *c = a + b; Let s take a look at main()

34 Addition on the Device: main() int main(void) { int a, b, c; int *d_c; int size = sizeof(int); // host copies of a, b, c // device copy of c // Allocate space for device copy of c cudamalloc((void **)&d_c, size); // Setup input values a = 2; b = 7;

35 Addition on the Device: main() // Launch add() kernel on GPU add<<<1,1>>>(a, b, d_c); // Copy result back to host cudamemcpy(&c, d_c, size, cudamemcpydevicetohost); } // Cleanup cudafree(d_a); cudafree(d_b); cudafree(d_c); return 0;

36 Unified Memory Simplifies Data Management int main(void) { int a, b; int *c; int size = sizeof(int); // Allocate space for device copy of c cudamallocmanged(&c, size); // Setup input values a = 2; b = 7; // Launch add() kernel on GPU add<<<1,1>>>(a, b, c); //use result on host, no explicit copy needed! cudadevicesynchronize(); printf( Result is %d\n, d); // Cleanup cudafree(c); return 0; }

37 Review Difference between host and device Host CPU Device GPU Using global to declare a function as device code (kernel) Executes on the device Called from the host Passing parameters from host code to a device kernel

38 RUNNING IN PARALLEL

39 Moving to Parallel GPU computing is about massive parallelism So how do we run code in parallel on the device? add<<< 1, 1 >>>(); add<<< 1, N >>>(); Instead of executing add() once, execute N threads in parallel

40 CUDA Threads With add() running in parallel we can do vector addition Terminology: a block can be split into parallel threads Let s change add() to use parallel threads instead of a single thread global void add(int *a, int *b, int *c) { c[blockidx.x] c[threadidx.x] = = a[blockidx.x] a[threadidx.x] + b[blockidx.x]; + b[threadidx.x]; } We use threadidx.x Let s have a look at main()

41 Unified Memory #define N 512 int main(void) { int *a, *b, *c; int size = N * sizeof(int); // Allocate space for N-vectors a, b, c cudamallocmanaged(&a, size); cudamallocmanaged(&b, size); cudamallocmanged(&c, size); // Setup input values on host fillrandom(n, a, b); // Launch one add() kernel block on GPU with N threads add<<< 1, N >>>(a, b, c); //use result on host, no explicit copy needed cudadevicesynchronize(); printvector(n, c); // Cleanup cudafree(a); cudafree(b); cudafree(c); return 0; }

42 Review Launching parallel kernels Kernels are launched with a given grid (1/2/3D index space) and block (1/2/3D index subspace) through the <<<grid, block>>> notation Launch a N blocks with M threads with add<<<n,m>>>( ); Use blockidx.x to access block index, threadidx.x to access thread index However, the number of threads per block is very limited (max 1024) Also, a block is always executed on a single streaming multiprocessor (SM), the other ~14 (on K40) SMs are unused The grid allows a much bigger index space to fill the GPU

43 WELCOME TO THE GRID

44 Indexing Arrays with Blocks and Threads A thread calculates its position in the 1/2/3D grid using blockidx.x and threadidx.x Consider indexing an array with one element per thread (8 threads/block) threadidx.x threadidx.x threadidx.x threadidx.x blockidx.x = 0 blockidx.x = 1 blockidx.x = 2 blockidx.x = 3 With M threads per block, a unique index for each thread is given by: int index = blockidx.x * M + threadidx.x;

45 Indexing Arrays: Example Which thread will operate on the red element? M = 8 threadidx.x = blockidx.x = 2 int index = blockidx.x * M + threadidx.x; = 2 * 8 + 5; = 21;

46 Vector Addition with Blocks and Threads Use the built-in variable blockdim.x for threads per block int index = blockidx.x * blockdim.x + threadidx.x; Combined version of add() to use parallel threads and parallel blocks global void add(int *a, int *b, int *c) { int index = blockidx.x * blockdim.x + threadidx.x; c[index] = a[index] + b[index]; } What changes need to be made in main()?

47 Unified Memory #define N (2048*2048) #define THREADS_PER_BLOCK 512 int main(void) { int *a, *b, *c; int size = N * sizeof(int); // Allocate space for device copies of a, b, c cudamallocmanaged(&a, size); cudamallocmanaged(&b, size); cudamallocmanged(&c, size); // Setup input values on host fillrandom(n, a, b); // Launch add() kernel on GPU with N blocks add<<<n/threads_per_block,threads_per_block>>>(a, b, c); //use result on host cudadevicesynchronize(); printvector(n, c); // Cleanup cudafree(a); cudafree(b); cudafree(c); return 0; }

48 Handling Arbitrary Vector Sizes Typical problems are not even multiples of blockdim.x Avoid accessing beyond the end of the arrays: global void add(int *a, int *b, int *c, int n) { int index = threadidx.x + blockidx.x * blockdim.x; if (index < n) c[index] = a[index] + b[index]; } Update the kernel launch: add<<<(n + M-1) / M,M>>>(d_a, d_b, d_c, N);

49 CONCEPTS Heterogeneous Computing Blocks Threads Indexing Shared memory syncthreads() Asynchronous operation ADVANCED CONCEPTS: COOPERATING THREADS Handling errors Managing devices

50 1D Stencil Consider applying a 1D stencil to a 1D array of elements Each output element is the sum of input elements within a radius If radius is 3, then each output element is the sum of 7 input elements: in out

51 Implementing Within a Block Each thread processes one output element blockdim.x elements per block Input elements are read several times With radius 3, each input element is read seven times in radius radius out Thread Thread Thread Thread Thread Thread Thread Thread Thread

52 Sharing Data Between Threads Terminology: within a block, threads share data via shared memory Extremely fast on-chip memory By opposition to device memory, referred to as global memory Like a user-managed cache Declare using shared, allocated per block Data is not visible to threads in other blocks

53 Implementing With Shared Memory Cache data in shared memory Read (blockdim.x + 2 * radius) input elements from global memory to shared memory Compute blockdim.x output elements Write blockdim.x output elements to global memory Each block needs a halo of radius elements at each boundary in out halo on left halo on right blockdim.x output elements

54 Stencil Kernel (1 of 2) global void stencil_1d(int *in, int *out) { shared int temp[block_size + 2 * RADIUS]; int gindex = threadidx.x + blockidx.x * blockdim.x; int lindex = threadidx.x + RADIUS; // Read input elements into shared memory temp[lindex] = in[gindex]; if (threadidx.x < RADIUS) { temp[lindex - RADIUS] = in[gindex - RADIUS]; } temp[lindex + BLOCK_SIZE] = in[gindex + BLOCK_SIZE];

55 Stencil Kernel (2 of 2) // Apply the stencil int result = 0; for (int offset = -RADIUS ; offset <= RADIUS ; offset++) result += temp[lindex + offset]; } // Store the result out[gindex] = result;

56 Data Race! The stencil example will not work Suppose thread 15 reads the halo before thread 0 has fetched it... temp[lindex] = in[gindex]; if (threadidx.x < RADIUS) { } temp[lindex RADIUS] = in[gindex RADIUS]; temp[lindex + BLOCK_SIZE] = in[gindex + BLOCK_SIZE]; int result = 0; for (int offset = -RADIUS ; offset <= RADIUS ; offset++)... result += temp[lindex + offset]; Store at temp[18] Load from temp[19] Skipped since threadid.x > RADIUS

57 syncthreads() void syncthreads(); Synchronizes all threads within a block Used to prevent RAW / WAR / WAW hazards All threads must reach the barrier In conditional code, the condition must be uniform across the block

58 Stencil Kernel global void stencil_1d(int *in, int *out) { shared int temp[block_size + 2 * RADIUS]; int gindex = threadidx.x + blockidx.x * blockdim.x; int lindex = threadidx.x + radius; // Read input elements into shared memory temp[lindex] = in[gindex]; if (threadidx.x < RADIUS) { temp[lindex RADIUS] = in[gindex RADIUS]; } temp[lindex + BLOCK_SIZE] = in[gindex + BLOCK_SIZE]; // Synchronize (ensure all the data is available) syncthreads();

59 Stencil Kernel // Apply the stencil int result = 0; for (int offset = -RADIUS ; offset <= RADIUS ; offset++) result += temp[lindex + offset]; } // Store the result out[gindex] = result;

60 Review (1 of 2) Launching parallel threads Launch N blocks with M threads per block with kernel<<<n,m>>>( ); Use blockidx.x to access block index within grid Use threadidx.x to access thread index within block Assign elements to threads: int index = blockidx.x * blockdim.x + threadidx.x;

61 Review (2 of 2) Use shared to declare a variable/array in shared memory Data is shared between threads in a block Not visible to threads in other blocks Use syncthreads() as a barrier Use to prevent data hazards

62 CONCEPTS Heterogeneous Computing Blocks Threads Indexing Shared memory syncthreads() Asynchronous operation MANAGING THE DEVICE Handling errors Managing devices

63 Coordinating Host & Device Kernel launches are asynchronous Control returns to the CPU immediately CPU needs to synchronize before consuming the results cudamemcpy() cudamemcpyasync() Blocks the CPU until the copy is complete Copy begins when all preceding CUDA calls have completed Asynchronous, does not block the CPU cudadevicesynchronize() Blocks the CPU until all preceding CUDA calls have completed

64 Reporting Errors All CUDA API calls return an error code (cudaerror_t) Error in the API call itself OR Error in an earlier asynchronous operation (e.g. kernel) Get the error code for the last error: cudaerror_t cudagetlasterror(void) Get a string to describe the error: char *cudageterrorstring(cudaerror_t) if(cudagetlasterror()!= cudasuccess) printf("%s\n", cudageterrorstring(cudagetlasterror()));

65 Device Management Application can query and select GPUs cudagetdevicecount(int *count) cudasetdevice(int device) cudagetdevice(int *device) cudagetdeviceproperties(cudadeviceprop *prop, int device) Multiple host threads can share a device A single host thread can manage multiple devices cudasetdevice(i) to select current device cudamemcpy( ) for peer-to-peer copies

66 Next Steps We just touched the surface But you can already do a lot of useful things with what we learned Next topics: Thread cooperation via shared memory and syncthreads() Asynchronous Kernel launches and memory operations Device management (cudasetdevice, error handling ) Atomic operations Launching kernels from the GPU MPI

67 Links and References These languages are supported on all CUDA-capable GPUs. CUDA C/C++ and Fortran Parallel Forall blog Thrust C++ Template Library Accelerated Libraries CUDA Programming Guide PyCUDA (Python) GTC On Demand gtcnew/on-demand-gtc.php

68 LUNCH