Skip to content

Latest commit

 

History

History
73 lines (62 loc) · 2.46 KB

cuda_cheatsheet.md

File metadata and controls

73 lines (62 loc) · 2.46 KB
  • cudaMalloc() - Allocate memory on the device
  • cudaMemcpy() - Copy memory from host to device and vice versa
  • Writing a kernel:
__global__ void add(int *a, int *b, int *c, int n) {
    int i = threadIdx.x + blockIdx.x * blockDim.x;
    if(i < n)
        c[i] = a[i] + b[i];
}
  • Launching a kernel:
#define TOTAL_THREADS (2048*2048)
#define THREADS_PER_BLOCK 512
// Allocate host memory
// Allocate device memory
// Copy host memory to device memory
add<<<TOTAL_THREADS/THREADS_PER_BLOCK, THREADS_PER_BLOCK>>>(d_a, d_b, d_c);
// Copy device memory to host memory

  • Memory types:

    • global memory
    • shared memory: on-chip memory. low latency, high bandwidth. Shared between threads in a block. Declared as __shared__ and user managed.
    • Typically > 48KB shared memory per SM.

  • __syncthreads() - Synchronize all threads in a block

  • 1D stencil example:

__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];
    }
    
    __syncthreads();

    int result = 0;
    for(int offset = -RADIUS; offset <= RADIUS; offset++)
        result += in[lindex + offset];
    
    out[gindex] = result;
}
  • Execution Model:
    • Threads can be mapped to functional units in SM.
    • 32 threads are grouped into a warp.
    • A warp is the smallest unit of execution. All threads in a warp execute the same instruction.
    • A block is a collection of threads and warps.
    • All the threads in a block are executed on the same SM.
    • Typically there are 64 warps per SM.
    • A block contains 32 * 64 = 2048 threads.
    • SM can concurrently execute 16-32 thread blocks.
  • GPUs are in order processors. They can execute multiple warps in parallel.
  • Memory reads doesn't stall execution. Threads stalls when operands are not ready.
  • Global memory latency 100-400 cycles.
  • Each SM has it's own L1 cache and L2 cache is shared by all SMs.
  • Cache line size is 128 bytes.
  • Loads can be caching or non-caching.
    • Caching loads: L1 cache hit -> L2 cache hit -> global memory
    • Non-caching loads: L1 cache miss -> L2 cache hit -> global memory
  • Stores invalidate L1 and write back to L2.