CUDA

Purpose

Guide agents through NVIDIA CUDA C/C++ development: kernel launch configuration, the memory hierarchy from registers through global memory, asynchronous execution with streams, nvcc compilation flags, Thrust library usage, and diagnosing common performance pitfalls like warp divergence and uncoalesced memory access.

When to Use

• Writing or optimizing a CUDA kernel for matrix multiply, reduction, or stencil operations
• Choosing block/grid dimensions and estimating occupancy
• Debugging slow kernels due to memory access patterns or low occupancy
• Setting up multi-stream pipelines with async cudaMemcpyAsync
• Compiling with nvcc and selecting architecture flags (-gencode)
• Using Thrust for parallel primitives instead of hand-written kernels

Workflow

1. Minimal kernel and launch

CUDA organizes work as threads grouped into blocks, blocks grouped into a grid.

Thread hierarchy
├── grid (1D/2D/3D)
│   └── block (1D/2D/3D, max 1024 threads)
│       └── thread (threadIdx, blockIdx, blockDim, gridDim)

// vector_add.cu
#include <cuda_runtime.h>
#include <stdio.h>

__global__ void vector_add(const float *a, const float *b, float *c, int n) {
    int i = blockIdx.x * blockDim.x + threadIdx.x;
    if (i < n)
        c[i] = a[i] + b[i];
}

int main(void) {
    const int n = 1 << 20;
    size_t bytes = n * sizeof(float);
    float *h_a, *h_b, *h_c, *d_a, *d_b, *d_c;

    cudaMalloc(&d_a, bytes);
    cudaMalloc(&d_b, bytes);
    cudaMalloc(&d_c, bytes);
    // ... host init and cudaMemcpy H2D ...

    int threads = 256;
    int blocks = (n + threads - 1) / threads;
    vector_add<<<blocks, threads>>>(d_a, d_b, d_c, n);
    cudaDeviceSynchronize();

    cudaMemcpy(h_c, d_c, bytes, cudaMemcpyDeviceToHost);
    cudaFree(d_a); cudaFree(d_b); cudaFree(d_c);
    return 0;
}

2. Memory hierarchy

Shared memory example (matrix tile):

#define TILE 16

__global__ void matmul_tiled(const float *A, const float *B, float *C, int N) {
    __shared__ float As[TILE][TILE];
    __shared__ float Bs[TILE][TILE];

    int row = blockIdx.y * TILE + threadIdx.y;
    int col = blockIdx.x * TILE + threadIdx.x;
    float sum = 0.0f;

    for (int t = 0; t < (N + TILE - 1) / TILE; t++) {
        As[threadIdx.y][threadIdx.x] = (row < N && t * TILE + threadIdx.x < N)
            ? A[row * N + t * TILE + threadIdx.x] : 0.0f;
        Bs[threadIdx.y][threadIdx.x] = (col < N && t * TILE + threadIdx.y < N)
            ? B[(t * TILE + threadIdx.y) * N + col] : 0.0f;
        __syncthreads();

        for (int k = 0; k < TILE; k++)
            sum += As[threadIdx.y][k] * Bs[k][threadIdx.x];
        __syncthreads();
    }
    if (row < N && col < N)
        C[row * N + col] = sum;
}

3. Streams and async copies

cudaStream_t stream1, stream2;
cudaStreamCreate(&stream1);
cudaStreamCreate(&stream2);

cudaMemcpyAsync(d_a, h_a, bytes, cudaMemcpyHostToDevice, stream1);
cudaMemcpyAsync(d_b, h_b, bytes, cudaMemcpyHostToDevice, stream2);
vector_add<<<blocks, threads, 0, stream1>>>(d_a, d_b, d_c, n);
cudaMemcpyAsync(h_c, d_c, bytes, cudaMemcpyDeviceToHost, stream1);
cudaStreamSynchronize(stream1);

Pinned host memory (cudaMallocHost) enables true async DMA overlap with kernel execution.

4. nvcc compilation

# Single architecture (local GPU)
nvcc -O3 -arch=sm_80 -o prog vector_add.cu

# Fat binary for multiple GPUs
nvcc -O3 \
  -gencode arch=compute_80,code=sm_80 \
  -gencode arch=compute_90,code=sm_90 \
  -o prog vector_add.cu

# Debug symbols for cuda-gdb
nvcc -G -g -O0 -arch=sm_80 -o prog_debug vector_add.cu

# Show PTX/SASS
nvcc -arch=sm_80 -ptx vector_add.cu
nvcc -arch=sm_80 -cubin vector_add.cu
cuobjdump -sass prog

Common flags:

5. Occupancy estimation

# CUDA Occupancy Calculator (spreadsheet) or programmatic:
./occupancy_tool --kernel vector_add --block-size 256 --regs 16 --smem 0

Decision tree:

Kernel slow?
├── Low occupancy (< 25%) → reduce registers, shared mem, or block size
├── Memory-bound → check coalescing, use shared memory tiling
└── Compute-bound → increase arithmetic intensity, use tensor cores

Use cudaOccupancyMaxActiveBlocksPerMultiprocessor API or Nsight Compute sm__warps_active.avg.pct_of_peak_sustained_active metric.

6. Thrust basics

#include <thrust/device_vector.h>
#include <thrust/sort.h>
#include <thrust/reduce.h>

thrust::device_vector<int> d_vec(1000000);
thrust::sort(d_vec.begin(), d_vec.end());
int sum = thrust::reduce(d_vec.begin(), d_vec.end());

Thrust handles temporary storage and kernel launches internally. Prefer Thrust for sort/scan/reduce; write custom kernels for domain-specific fused operations.

7. Common pitfalls

Warp divergence: Threads in a warp (32) execute in SIMT lockstep. Branching on threadIdx causes serialization.

// Bad: divergent branch
if (threadIdx.x % 2 == 0) { heavy_a(); } else { heavy_b(); }

// Better: separate kernels or predication

Uncoalesced access: Consecutive threads should access consecutive addresses.

// Bad: strided access
float val = data[threadIdx.x * stride];

// Good: coalesced
float val = data[blockIdx.x * blockDim.x + threadIdx.x];

Common Problems

Related Skills

• skills/gpu/cuda-profiling — Nsight Systems/Compute for kernel performance diagnosis
• skills/gpu/cuda-debugging — cuda-gdb and compute-sanitizer for correctness
• skills/gpu/gpu-memory-model — SIMT execution, coalescing rules, bank conflicts
• skills/gpu/hip-rocm — porting CUDA kernels to AMD HIP
• skills/gpu/triton-lang — Python DSL alternative for custom kernels
• skills/hpc/openmp — host-side parallelism alongside CUDA