obsidian/study/CUDA notes.md

Thread Hierarchy

thread, block, (cluster), grid

for a two-dimensional block of size (Dx, Dy), the thread ID of a thread of index (x, y) is (x + y * Dx); for a three-dimensional block of size (Dx, Dy, Dz), the thread ID of a thread of index (x, y, z) is (x + y * Dx + z * Dx * Dy)

SIMT architecture (single-instruction multi-thread)

32 threads as a warp 一个 warp 内 single instruction，不同 thread 若执行的指令流一致，则并行，否则分别执行，因此尽可能保证一个 wrap 内的 thread 执行相同的指令流。Actually in nowadays:

Starting with the NVIDIA Volta architecture, Independent Thread Scheduling allows full concurrency between threads, regardless of warp. threads can now diverge and reconverge at sub-warp granularity.

You can test it by following code:

#include <chrono>
#include <iostream>

#define T 1000000

__global__ void f(float *x) {
  int i = blockIdx.x * blockDim.x + threadIdx.x;
  for (int t = 0; t < T; t++) {
	if (i < (1 << 10)) {
	  x[i] *= 1.1;
	} else {
	  x[i] += 1;
	}

	// if (i % 2 == 0) {
	//   x[i] *= 1.1;
	// } else {
	//   x[i] += 1;
	// }

	if (x[i] > 1e6) {
	  x[i] = 1.0;
	}
  }
}

int main(void) {
  int N = 1 << 20;
  float *x, *d_x;
  x = (float *)malloc(N * sizeof(float));

  cudaMalloc(&d_x, N * sizeof(float));
  for (int i = 0; i < N; i++) {
	x[i] = 1.0f;
  }
  cudaMemcpy(d_x, x, N * sizeof(float), cudaMemcpyHostToDevice);

  auto start = std::chrono::steady_clock::now();
  f<<<N / 256, 256>>>(d_x);
  auto end = std::chrono::steady_clock::now();
  auto elapse = end - start;
  std::cout << "time used: " << elapse.count() << " ns" << '\n';

  cudaMemcpy(x, d_x, N * sizeof(float), cudaMemcpyDeviceToHost);

  cudaFree(d_x);
  free(x);
}

Summary