CUDA float2合并

问题描述

在CUDA中使用float2数据类型时，合并读取时遇到问题。

I'm having trouble coalescing reads when using the float2 datatype in CUDA.

我试图做一个简单的示例在可视化探查器中运行，但是它总是返回非终止读取。如果有人能对此有所阐明，我将非常感谢。

I've tried to make a simple example to run in the visual profiler but it always returns noncoalesced reads. If anyone could shed some light on this I would be really grateful, thanks.

#include <stdio.h>
#include <cuda_runtime_api.h>

__global__ void kernel(float2 *in, float2 *out) {
        int idx=blockIdx.x*blockDim.x+threadIdx.x;
        float2 d=in[idx];
        d.x = 100.f;

        out[idx] = d;
}

int main() {
  const int dataSize=32;
  float2 *in;
  cudaMalloc((void**)&in,dataSize*sizeof(float2));

  float2 *out;
  cudaMalloc((void**)&out,dataSize*sizeof(float2));
  kernel<<<1,32>>>(in,out);
  return 0;
}

推荐答案

我在NVIDIA论坛。事实证明，在调试模式下未对加载向量进行优化。

I asked this question on the NVIDIA forums. Turns out loading vectors is not optimized in debug mode. Forums

这篇关于CUDA float2合并的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持！