Introduction to CUDA Programming CUDA Programming Introduction Andreas Moshovos Winter 2009 Some slides/material from: UIUC course by Wen-Mei Hwu and David Kirk UCSB course by Andrea Di Blas Universitat Jena by Waqar Saleem NVIDIA by Simon Green

Programmer’s view GPU as a co-processor CPU Memory GPU GPU Memory 1GB on our systems 3GB/s – 8GB.s 6.4GB/sec – 31.92GB/sec 8B per transfer 141GB/sec

Target Applications int a[N]; // N is large for all elements of a compute a[i] = a[i] * fade Lots of independent computations –CUDA thread need not be independent

Programmer’s View of the GPU GPU: a compute device that: –Is a coprocessor to the CPU or host –Has its own DRAM (device memory) –Runs many threads in parallel Data-parallel portions of an application are executed on the device as kernels which run in parallel on many threads

Why are threads useful? Concurrency: –Do multiple things in parallel –Put more hardware  Get higher performance Needs more functional units

Why are threads useful #2 – Tolerating stalls Often a thread stalls, e.g., memory access Multiplex the same functional unit Get more performance at a fraction of the cost

GPU vs. CPU Threads GPU threads are extremely lightweight Very little creation overhead In the order of microseconds GPU needs 1000s of threads for full efficiency Multi-core CPU needs only a few

Execution Timeline time 1. Copy to GPU mem 2. Launch GPU Kernel GPU / Device 2’. Synchronize with GPU 3. Copy from GPU mem CPU / Host

GBT: Grids of Blocks of Threads Why? Realities of integrated circuits: need to cluster computation and storage to achieve high speeds

Grids of Blocks of Threads: Dimension Limits Grid of Blocks 1D or 2D –Max x: –Max y: Block of Threads: 1D, 2D, or 3D –Max number of threads: 512 –Max x: 512 –Max y: 512 –Max z: 64 Limits apply to Compute Capability 1.0, 1.1, 1.2, and 1.3 –GTX280 = 1.3

Block and Thread IDs Threads and blocks have IDs –So each thread can decide what data to work on –Block ID: 1D or 2D –Thread ID: 1D, 2D, or 3D Simplifies memory addressing when processing multidimensional data Device Grid 1 Block (0, 0) Block (1, 0) Block (2, 0) Block (0, 1) Block (1, 1) Block (2, 1) Block (1, 1) Thread (0, 1) Thread (1, 1) Thread (2, 1) Thread (3, 1) Thread (4, 1) Thread (0, 2) Thread (1, 2) Thread (2, 2) Thread (3, 2) Thread (4, 2) Thread (0, 0) Thread (1, 0) Thread (2, 0) Thread (3, 0) Thread (4, 0) IDs and dimensions are easily accessible through predefined “variables”, e.g., blockDim.x and threadIdx.x

Thread Batching A kernel is executed as a grid of thread blocks –All threads share data memory space A thread block: threads that can cooperate with each other by: –Synchronizing their execution For hazard-free shared memory accesses –Efficiently sharing data through a low latency shared memory Two threads from two different blocks cannot cooperate Host Kernel 1 Kernel 2 Device Grid 1 Block (0, 0) Block (1, 0) Block (2, 0) Block (0, 1) Block (1, 1) Block (2, 1) Grid 2 Block (1, 1) Thread (0, 1) Thread (1, 1) Thread (2, 1) Thread (3, 1) Thread (4, 1) Thread (0, 2) Thread (1, 2) Thread (2, 2) Thread (3, 2) Thread (4, 2) Thread (0, 0) Thread (1, 0) Thread (2, 0) Thread (3, 0) Thread (4, 0)

Thread Coordination Overview Race-free access to data

Programmer’s view: Memory Model

Programmer’s View: Memory Detail – Thread and Host Each thread can: –R/W per-thread registers –R/W per-thread local memory –R/W per-block shared memory –R/W per-grid global memory –Read only per-grid constant memory –Read only per-grid texture memory The host can R/W: global, constant, and texture memories

Memory Model: Global, Constant, and Texture Memories Global memory –Main means of communicating R/W Data between host and device –Contents visible to all threads Texture and Constant Memories –Constants initialized by host –Contents visible to all threads

Memory Model Summary MemoryLocationCachedAccessScope Localoff-chipNoR/Wthread Sharedon-chipN/AR/Wall threads in a block Globaloff-chipNoR/Wall threads + host Constantoff-chipYesROall threads + host Textureoff-chipYesROall threads + host

Execution Model: Ordering Execution order is undefined Do not assume and use: block 0 executes before block 1 Thread 10 executes before thread 20 And any other ordering even if you can observe it

Execution Model Summary Grid of blocks of threads –1D/2D grid of blocks –1D/2D/3D blocks of threads All blocks are identical: –same structure and # of threads Block execution order is undefined Same block threads: –can synchronize and share data fast (shared memory) Threads from different blocks: –Cannot cooperate –Communication through global memory Threads and Blocks have IDs –Simplifies data indexing –Can be 1D, 2D, or 3D (threads) Blocks do not migrate: execute on the same processor Several blocks may run over the same processor

CUDA Software Architecture cuda…() cu…() e.g., fft()

Reasoning about CUDA call ordering GPU communication via cuda…() calls and kernel invocations –cudaMalloc, cudaMemCpy, Asynchronous from the CPU’s perspective –CPU places a request in a “CUDA” queue –requests are handled in-order Streams allow for multiple queues –More on this much later one

CUDA API: Example int a[N]; for (i =0; i < N; i++) a[i] = a[i] + x; 1.Allocate CPU Data Structure 2.Initialize Data on CPU 3.Allocate GPU Data Structure 4.Copy Data from CPU to GPU 5.Define Execution Configuration 6.Run Kernel 7.CPU synchronizes with GPU 8.Copy Data from GPU to CPU 9.De-allocate GPU and CPU memory

1. Allocate CPU Data float *ha; main (int argc, char *argv[]) { int N = atoi (argv[1]); ha = (float *) malloc (sizeof (float) * N);... } Pinned memory allocation results in faster CPU to/from GPU copies More on this later cudaMallocHost (…)

2. Initialize CPU Data float *ha; int i; for (i = 0; i < N; i++) ha[i] = i;

3. Allocate GPU Data float *da; cudaMalloc ((void **) &da, sizeof (float) * N); Notice: no assignment side –NOT: da = cudaMalloc (…) Assignment is done internally: –That why we pass &da Allocated space in Global Memory

GPU Memory Allocation The host manages GPU memory allocation: –cudaMalloc (void **ptr, size_t nbytes) –Must explicitly cast to ( void **) cudaMalloc ((void **) &da, sizeof (float) * N); –cudaFree (void *ptr); cudaFree (da); –cudaMemset (void *ptr, int value, size_t nbytes); –cudaMemset (da, 0, N * sizeof (int)); Check the CUDA Reference Manual

4. Copy Initialized CPU data to GPU float *da; float *ha; cudaMemCpy ((void *) da, // DESTINATION (void *) ha, // SOURCE sizeof (float) * N, // #bytes cudaMemcpyHostToDevice); // DIRECTION

Host/Device Data Transfers The host initiates all transfers: cudaMemcpy(void *dst, void *src, size_t nbytes, enum cudaMemcpyKind direction) Asynchronous from the CPU’s perspective –CPU thread continues In-order processing with other CUDA requests enum cudaMemcpyKind –cudaMemcpyHostToDevice –cudaMemcpyDeviceToHost –cudaMemcpyDeviceToDevice

5. Define Execution Configuration How many blocks and threads/block int threads_block = 64; int blocks = N / threads_block; if (blocks % N != 0) blocks += 1; Alternatively: blocks = (N + threads_block – 1) / threads_block;

6. Launch Kernel & 7. CPU/GPU Synchronization Instructs the GPU to launch blocks x thread_blocks threads: darradd > (da, 10f, N); cudaThreadSynchronize (); darradd: kernel name >> execution configuration –More on this soon (da, x, N): arguments –256 – 8 byte limit / No variable arguments

CPU/GPU Synchronization CPU does not block on cuda…() calls –Kernel/requests are queued and processed in-order –Control returns to CPU immediately Good if there is other work to be done –e.g., preparing for the next kernel invocation Eventually, CPU must know when GPU is done Then it can safely copy the GPU results cudaThreadSynchronize () –Block CPU until all preceding cuda…() and kernel requests have completed

8. Copy data from GPU to CPU & 9. DeAllocate Memory float *da; float *ha; cudaMemCpy ((void *) ha, // DESTINATION (void *) da, // SOURCE sizeof (float) * N, // #bytes cudaMemcpyDeviceToHost); // DIRECTION cudaFree (da); // display or process results here free (ha);

The GPU Kernel __global__ darradd (float *da, float x, int N) { int i = blockIdx.x * blockDim.x + threadIdx.x; if (i < N) da[i] = da[i] + x; } BlockIdx: Unique Block ID. –Numerically asceding: 0, 1, … ThreadIdx: Unique per Block Index –0, 1, … –Per Block BlockDim: Dimensions of Block –BlockDim.x, BlockDim.y, BlockDim.z –Unused dimensions default to 0

Array Index Calculation Example int i = blockIdx.x * blockDim.x + threadIdx.x; a[0]a[63]a[64]a[127]a[128]a[255]a[256] blockIdx.x 0blockIdx.x 1blockIdx.x 2 threadIdx.x 0 threadIdx.x 63 threadIdx.x 0 threadIdx.x 63 threadIdx.x 0 threadIdx.x 63 threadIdx.x 0 i = 0i = 63i = 64i = 127i = 128i = 255 i = 256 Assuming blockDim.x = 64

Generic Unique Thread and Block Index Calculations #1 1D Grid / 1D Blocks: UniqueBlockIndex = blockIdx.x; UniqueThreadIndex = blockIdx.x * blockDim.x + threadIdx.x; 1D Grid / 2D Blocks: UniqueBlockIndex = blockIdx.x; UniqueThreadIndex = blockIdx.x * blockDim.x * blockDim.y + threadIdx.y * blockDim.x + threadIdx.x; 1D Grid / 3D Blocks: UniqueBockIndex = blockIdx.x; UniqueThreadIndex = blockIdx.x * blockDim.x * blockDim.y * blockDim.z + threadIdx.z * blockDim.y * blockDim.x + threadIdx.y * blockDim.x + threadIdx.x; Source:

Generic Unique Thread and Block Index Calculations #2 2D Grid / 1D Blocks: UniqueBlockIndex = blockIdx.y * gridDim.x + blockIdx.x; UniqueThreadIndex = UniqueBlockIndex * blockDim.x + threadIdx.x; 2D Grid / 2D Blocks: UniqueBlockIndex = blockIdx.y * gridDim.x + blockIdx.x; UniqueThreadIndex =UniqueBlockIndex * blockDim.y * blockDim.x + threadIdx.y * blockDim.x + threadIdx.x; 2D Grid / 3D Blocks: UniqueBlockIndex = blockIdx.y * gridDim.x + blockIdx.x; UniqueThreadIndex = UniqueBlockIndex * blockDim.z * blockDim.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.z + threadIdx.y * blockDim.x + threadIdx.x; UniqueThreadIndex means unique per grid.

CUDA Function Declarations __global__ defines a kernel function –Must return void –Can only call __device__ functions __device__ and __host__ can be used together Executed on the: Only callable from the: __device__ float DeviceFunc() device __global__ void KernelFunc() devicehost __host__ float HostFunc() host

__device__ Example Add x to a[i] multiple times __device__ float addmany (float a, float b, int count) { while (count--) a += b; return a; } __global__ darradd (float *da, float x, int N) { int i = blockIdx.x * blockDim.x + threadIdx.x; if (i < N) da[i] = addmany (da[i], x, 10); }

Kernel and Device Function Restrictions __device__ functions cannot have their address taken –e.g., f = &addmany; *f(…); For functions executed on the device: –No recursion darradd (…) { darradd (…) } –No static variable declarations inside the function darradd (…) { static int canthavethis; } –No variable number of arguments e.g., something like printf (…)

Predefined Vector Datatypes Can be used both in host and in device code. –[u]char[1..4], [u]short[1..4], [u]int[1..4], [u]long[1..4], float[1..4] Structures accessed with.x,.y,.z,.w fields default constructors, “ make_TYPE (…)”: –float4 f4 = make_float4 (1f, 10f, 1.2f, 0.5f); dim3 –type built on uint3 –Used to specify dimensions –Default value is (1, 1, 1)

Execution Configuration Must specify when calling a __global__ function: >> where: –dim3 Dg : grid dimensions in blocks –dim3 Db : block dimensions in threads –size_t Ns : per block additional number of shared memory bytes to allocate optional, defaults to 0 more on this much later on –cudaStream_t S : request stream(queue) optional, default to 0. Compute capability >= 1.1

Built-in Variables dim3 gridDim –Number of blocks per grid, in 2D (.z always 1) uint3 blockIdx –Block ID, in 2D (blockIdx.z = 1 always) dim3 blockDim –Number of threads per block, in 3D uint3 threadIdx –Thread ID in block, in 3D

Execution Configuration Examples 1D grid / 1D blocks dim3 gd(1024) dim3 bd(64) akernel >>(...) gridDim.x = 1024, gridDim.y = 1, blockDim.x = 64, blockDim.y = 1, blockDim.z = 1 2D grid / 3D blocks dim3 gd(4, 128) dim3 bd(64, 16, 4) akernel >>(...) gridDim.x = 4, gridDim.y = 128, blockDim.x = 64, blockDim.y = 16, blockDim.z = 4

Error Handling Most cuda…() functions return a cudaError_t –If cudaSuccess : Request completed without a problem cudaGetLastError(): –returns the last error to the CPU –Use with cudaThreadSynchronize(): cudaError_t code; cudaThreadSynchronize (); code = cudaGetLastError (); char *cudaGetErrorString(cudaError_t code); –returns a human-readable description of the error code

Error Handling Utility Function void cudaDie (const char *msg) { cudaError_t err; cudaThreadSynchronize (); err = cudaGetLastError(); if (err == cudaSuccess) return; fprintf (stderr, "CUDA error: %s: %s.\n", msg, cudaGetErrorString (err)); exit(EXIT_FAILURE); } adapted from:

Error Handling Macros CUDA_SAFE_CALL ( some cuda call ) CUDA_SAFE_CALL (cudaMemcpy (a_h, a_d, arr_size, cudaMemcpyDeviceToHost) ); Must define #define _DEBUG –No code emitted when undefined: Performance Use make dbg=1 under NVIDIA_CUDA_SDK

Measuring Time -- gettimeofday Unix-based: #include struct timeval start, end; gettimeofday (&start, NULL); WHAT WE ARE INTERESTED IN gettimeofday (&end, NULL); timeCpu = (float)(end.tv_sec - start.tv_sec); if (end.tv_usec < start.tv_usec) { timeCpu -= 1.0; timeCpu += (double)( end.tv_usec - start.tv_usec)/ ; } else timeCpu += (double)(end.tv_usec - start.tv_usec)/ ;

Using CUDA clock () clock_t clock (); Can be used in device code returns per multiprocessor counter which is incremented every clock cycle Sample at the beginning and end upper bound since threads are time-sliced uint start = clock();... compute (less than 3 sec).... uint end = clock(); if (end > start) time = end - start; else time = end + (0xffffffff - start) Look at the clock example under projects in SDK Using takes some effort –Every thread measures start and end –Then must find min start and max end –Accurate

Using cutTimer…() library calls #include unsigned int htimer; cutCreateTimer (&htimer); CudaThreadSynchronize (); cutStartTimer(htimer); WHAT WE ARE INTERESTED IN cudaThreadSynchronize (); cutStopTimer(htimer); printf (“time: %f\n", cutGetTimerValue(htimer));

Code Overview: Host side #include unsigned int htimer; float *ha, *da; main (int argc, char *argv[]) { int N = atoi (argv[1]); ha = (float *) malloc (sizeof (float) * N); for (int i = 0; i < N; i++) ha[i] = i; cutCreateTimer (&htimer); cudaMalloc ((void **) &da, sizeof (float) * N); cudaMemCpy ((void *) da, (void *) ha, sizeof (float) * N, cudaMemcpyHostToDevice); cudaThreadSynchronize (); cutStartTimer(htimer); blocks = (N + threads_block – 1) / threads_block; darradd > (da, 10f, N) cudaThreadSynchronize (); cutStopTimer(htimer); cudaMemCpy ((void *) ha, (void *) da, sizeof (float) * N, cudaMemcpyDeviceToHost); cudaFree (da); free (ha); printf (“processing time: %f\n", cutGetTimerValue(htimer)); }

Code Overview: Device Side __device__ float addmany (float a, float b, int count) { while (count--) a += b; return a; } __global__ darradd (float *da, float x, int N) { int i = blockIdx.x * blockDim.x + threadIdx.x; if (i < N) da[i] = addmany (da[i], x, 10); }

Variable Declarations – Will revisit next time __device__ –stored in device memory (large, high latency, no cache) –Allocated with cudaMalloc (__device__qualifier implied) –accessible by all threads –lifetime: application __constant__ –same as __device__, but cached and read-only by GPU –written by CPU via cudaMemcpyToSymbol(...) call –lifetime: application __shared__ –stored in on-chip shared memory (very low latency) –accessible by all threads in the same thread block –lifetime: kernel launch Unqualified variables: –scalars and built-in vector types are stored in registers –arrays of more than 4 elements or run-time indices stored in device memory

Measurement Methodology You will not get exactly the same time measurements every time –Other processes running / external events (e.g., network activity) –Cannot control –“Non-determinism” Must take sufficient samples –say 10 or more –There is theory on what the number of samples must be Measure average Will discuss this next time or will provide a handout online

Handling Large Input Data Sets – 1D Example Recall gridDim.[xy] <= Host calls kernel multiple times: float *dac = da; // starting offset for current kernel while (n_blocks) { int bn = n_blocks; int elems; // array elements processed in this kernel if (bn > 65535) bn = 65535; elems = bn * block_size; darradd >> (dac, 10.0f, elems); n_blocks -= bn; dac += elems; }