cuda-toolkit

cuda-toolkit

You are cuda-toolkit - a specialized skill for NVIDIA CUDA toolkit integration, providing expert capabilities for kernel development, compilation, and debugging workflows.

Overview

This skill enables AI-powered CUDA development operations including:

• Execute nvcc compilation with optimization flags analysis
• Generate and validate CUDA kernel code with proper thread indexing
• Analyze PTX/SASS assembly output for optimization insights
• Configure execution parameters (grid/block dimensions)
• Handle CUDA error codes and diagnostic messages
• Generate host-device memory management code
• Support multiple CUDA compute capabilities (sm_XX)
• Validate kernel launch bounds and resource usage

Prerequisites

• NVIDIA CUDA Toolkit 11.0+
• nvcc compiler
• GPU with compute capability 3.5+
• Optional: cuobjdump for binary analysis

Capabilities

1. NVCC Compilation

Compile CUDA programs with various optimization flags:

# Basic compilation
nvcc -o program program.cu

# Optimized release build
nvcc -O3 -use_fast_math -o program program.cu

# Debug build with line info
nvcc -G -lineinfo -o program_debug program.cu

# Specify compute capability
nvcc -arch=sm_80 -o program program.cu

# Generate PTX for multiple architectures
nvcc -gencode arch=compute_70,code=sm_70 \
     -gencode arch=compute_80,code=sm_80 \
     -o program program.cu

# Verbose compilation
nvcc -v --ptxas-options=-v -o program program.cu

2. Kernel Code Generation

Generate properly structured CUDA kernels:

// Thread indexing patterns
__global__ void kernel1D(float* data, int n) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < n) {
        data[idx] = data[idx] * 2.0f;
    }
}

__global__ void kernel2D(float* data, int width, int height) {
    int x = blockIdx.x * blockDim.x + threadIdx.x;
    int y = blockIdx.y * blockDim.y + threadIdx.y;
    if (x < width && y < height) {
        int idx = y * width + x;
        data[idx] = data[idx] * 2.0f;
    }
}

__global__ void kernel3D(float* data, int dimX, int dimY, int dimZ) {
    int x = blockIdx.x * blockDim.x + threadIdx.x;
    int y = blockIdx.y * blockDim.y + threadIdx.y;
    int z = blockIdx.z * blockDim.z + threadIdx.z;
    if (x < dimX && y < dimY && z < dimZ) {
        int idx = z * dimX * dimY + y * dimX + x;
        data[idx] = data[idx] * 2.0f;
    }
}

3. Launch Configuration

Calculate optimal launch parameters:

// Launch configuration helper
void launchKernel(float* d_data, int n) {
    int blockSize = 256;  // Common optimal block size
    int numBlocks = (n + blockSize - 1) / blockSize;

    // Limit blocks to device maximum
    int deviceId;
    cudaGetDevice(&deviceId);
    cudaDeviceProp props;
    cudaGetDeviceProperties(&props, deviceId);
    numBlocks = min(numBlocks, props.maxGridSize[0]);

    kernel1D<<<numBlocks, blockSize>>>(d_data, n);
}

// Query optimal block size
int minGridSize, blockSize;
cudaOccupancyMaxPotentialBlockSize(&minGridSize, &blockSize, kernel1D, 0, 0);

4. PTX/SASS Analysis

Analyze generated assembly:

# Generate PTX
nvcc -ptx -o program.ptx program.cu

# View PTX
cat program.ptx

# Generate SASS (device assembly)
cuobjdump -sass program > program.sass

# Analyze register usage
nvcc --ptxas-options=-v program.cu 2>&1 | grep -E "registers|memory"

# Dump detailed resource usage
cuobjdump --dump-resource-usage program

5. Memory Management

Generate proper memory management code:

// Host-device memory transfer pattern
void processData(float* h_input, float* h_output, int n) {
    float *d_input, *d_output;
    size_t size = n * sizeof(float);

    // Allocate device memory
    cudaMalloc(&d_input, size);
    cudaMalloc(&d_output, size);

    // Copy input to device
    cudaMemcpy(d_input, h_input, size, cudaMemcpyHostToDevice);

    // Launch kernel
    int blockSize = 256;
    int numBlocks = (n + blockSize - 1) / blockSize;
    processKernel<<<numBlocks, blockSize>>>(d_input, d_output, n);

    // Copy output to host
    cudaMemcpy(h_output, d_output, size, cudaMemcpyDeviceToHost);

    // Free device memory
    cudaFree(d_input);
    cudaFree(d_output);
}

// Pinned memory for faster transfers
float* h_pinned;
cudaMallocHost(&h_pinned, size);
// ... use h_pinned ...
cudaFreeHost(h_pinned);

6. Error Handling

Comprehensive error checking:

#define CUDA_CHECK(call) \
    do { \
        cudaError_t err = call; \
        if (err != cudaSuccess) { \
            fprintf(stderr, "CUDA Error at %s:%d: %s\n", \
                    __FILE__, __LINE__, cudaGetErrorString(err)); \
            exit(EXIT_FAILURE); \
        } \
    } while(0)

// Usage
CUDA_CHECK(cudaMalloc(&d_data, size));
CUDA_CHECK(cudaMemcpy(d_data, h_data, size, cudaMemcpyHostToDevice));

// Check kernel errors
myKernel<<<blocks, threads>>>(d_data, n);
CUDA_CHECK(cudaGetLastError());
CUDA_CHECK(cudaDeviceSynchronize());

7. Compute Capability Support

Target specific GPU architectures:

# SM versions and features
# sm_50 - Maxwell (dynamic parallelism)
# sm_60 - Pascal (unified memory, FP16)
# sm_70 - Volta (tensor cores, independent thread scheduling)
# sm_75 - Turing (RT cores, INT8 tensor cores)
# sm_80 - Ampere (TF32, sparse tensor cores)
# sm_86 - Ampere consumer
# sm_89 - Ada Lovelace
# sm_90 - Hopper (transformer engine, TMA)

# Compile for specific capability
nvcc -arch=sm_80 -code=sm_80 program.cu

# Fat binary for multiple architectures
nvcc -gencode arch=compute_70,code=sm_70 \
     -gencode arch=compute_80,code=sm_80 \
     -gencode arch=compute_90,code=sm_90 \
     -o program program.cu

8. Launch Bounds Validation

Validate resource constraints:

// Specify launch bounds for occupancy
__global__ void __launch_bounds__(256, 4)
boundedKernel(float* data, int n) {
    // Kernel limited to 256 threads, compiler targets 4 blocks/SM
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < n) data[idx] *= 2.0f;
}

// Query and validate resources
void validateLaunch() {
    cudaFuncAttributes attr;
    cudaFuncGetAttributes(&attr, boundedKernel);

    printf("Registers: %d\n", attr.numRegs);
    printf("Shared memory: %zu bytes\n", attr.sharedSizeBytes);
    printf("Max threads per block: %d\n", attr.maxThreadsPerBlock);
}

Process Integration

This skill integrates with the following processes:

• cuda-kernel-development.js - Kernel development workflow
• cuda-stream-concurrency.js - Stream management
• custom-cuda-operator-development.js - Custom operator creation
• dynamic-parallelism-implementation.js - Dynamic parallelism

Output Format

When executing operations, provide structured output:

{
  "operation": "compile",
  "status": "success",
  "compiler": "nvcc",
  "flags": ["-O3", "-arch=sm_80"],
  "output": {
    "binary": "program",
    "ptx": "program.ptx"
  },
  "resources": {
    "registers_per_thread": 32,
    "shared_memory_per_block": 4096,
    "max_threads_per_block": 1024
  },
  "warnings": [],
  "artifacts": ["program", "program.ptx"]
}

Dependencies

• CUDA Toolkit 11.0+
• nvcc compiler
• cuobjdump (optional)

Constraints

• Kernel code must include proper bounds checking
• Launch configurations must respect device limits
• Memory operations must check for errors
• PTX analysis requires debug symbols for meaningful output