EBU6502_cloud_computing_notes/2-2-cuda.md

# CUDA

## Definition

- Compute Unified Device Architecture
- Is a general purpose **parallel** programming platform and model

## Scalability with CUDA

- GPU is has an array of Streaming Multiprocessors
- A program can be partitioned into blocks that execute independently
- Allow performance to scale with the number of GPU multiprocessors

## Architecture

### Kernel

- CUDA extends C by allow to define C functions, which is called a kernel
    - Kernel is executed in parallel, by N different CUDA threads, rather than
      only once in serial
    - Defined with specifier `__global__`
- Number of threads is given using `<<<NBlocks, NThreads>>>`, and used with
  `func_name<<<NBlocks, NThreads>>>(params);`
    - They take an integer or dim3, so it cal also be called Dimension of
      Blocks, or Dimension of Threads
- Thread number is given an ID, accessible by `threadIdx` variable
- Example:
    ```c
    // Add array A and B of size N together, store it in C
    #define N 2
    // Kernel definition
    __global__ void VecAdd(float* A, float* B, float* C) {
        // Each thread performs one pair wise addition
        int i = threadIdx.x;
        // The array is a float, meaning we need to allocate memory on GPU
        C[i] = A[i] + B[i];
    }
    int main() {
        ...
        // Kernel invocation with N threads
        VecAdd<<<1, N>>>(A, B, C);
        ...
    }
    ```

### Device code and Host code

- Compiler `nvcc` separates source code into host and device code
    - Device code (kernels): definition marked with `__global__`, called from
      host code (with `func<<<x, y>>>(a, b)`), and run on device (GPU)
    - Host code: normal C code, processed by system compiler like `gcc` or
      `clang`
- If the device code definitions are used, the compiled code execute on both CPU
  and GPU

### Thread hierarchy

- `threadIdx` is a 3 component vector
- Enables threads to be identified in 1-3 dimension blocks, which can form a
  **grid** of thread blocks
- Number of threads depend on the size of data being processed
- Features:
    - Maximum number of threads per block: because they reside on the same
      block, on the same processor core
    - A kernel can be executed by multiple equally shaped thread blocks, so the
      total number of threads is threads per block times blocks
- Example:
    ```c
    #define N 10
    // Kernel definition
    __global__ void MatAdd(float A[N][N], float B[N][N], float C[N][N]) {
        int i = threadIdx.x;
        int j = threadIdx.y;
        C[i][j] = A[i][j] + B[i][j];
    }
    int main() {
        ...
        // Kernel invocation with one block of N * N * 1 threads
        int numBlocks = 1;
        dim3 threadsPerBlock(N, N);
        MatAdd<<<numBlocks, threadsPerBlock>>>(A, B, C);
        ...
    }
    ```
- A complete example:

    ```c

    ```

### Block Hierarchy

- Blocks can be represented in 1-3 dimensional **grids**, size is dependent on
  the input data
- Accessing
    - Block index: `blockIdx`
    - Dimension: `blockDim`
- Example: multi block kernel:
    ```c
    #define N 16
    // Kernel definition
    __global__ void MatAdd(float A[N][N], float B[N][N],
    float C[N][N]) {
        int i = blockIdx.x * blockDim.x + threadIdx.x;
        int j = blockIdx.y * blockDim.y + threadIdx.y;
        if (i < N && j < N)
        C[i][j] = A[i][j] + B[i][j];
    }
    int main() {
        ...
        // Kernel invocation
        dim3 threadsPerBlock(16, 16);
        dim3 numBlocks(N / threadsPerBlock.x, N / threadsPerBlock.y);
        MatAdd<<<numBlocks, threadsPerBlock>>>(A, B, C);
        ...
    }
    ```

### Memory Hierarchy

- Thread: Registers, Local memory
- Block of threads: Shared memory
- All Blocks: Global memory

### GPU vs. CPU

- GPU has more stream processors, and each processor has much more ALUs
- TODO: see code in assets/code, and figure shit out, and index array

## Coordinating with CPU

- GPU kernel launch are **asynchronous**, while control returns to CPU
  immediately
- CPU need to **synchronize** before using the results
    - `cudaMemcpy()`: block and copy
    - `cudaMemcpyAsync()`: copy but don't block
    - `cudaDeviceSynchronize()`: Block the CPU until all CUDA calls before have
      completed