Compute shader中uint3与float乘法运算精度严重失准如何解决

问题表现 #

在Compute Shader中实现float类型与uint3类型的乘法运算时，计算结果存在远超浮点误差范围的精度偏差：

• 测试运算uint3(1, 2, 3) * 0.25
  • 预期结果：float3(0.25, 0.5, 0.75)
  • 实际输出：float3(0.3, 0.5, 0.8)
• 最小复现运算uint3(1, 1, 1) * 0.25
  • 预期结果：float3(0.25, 0.25, 0.25)
  • 实际输出：float3(0.3, 0.3, 0.3)
• 测试硬件：RTX 2070

问题定位 #

精度异常出现在以下代码行：

float3 node_min = MortonToXYZ(mid) * node_size; 

涉及参数说明：

• mid：当前线程对应的Morton码
• MortonToXYZ：Morton码解码函数，返回值为取值范围[0, +∞)的uint3类型
• node_size：取值范围[0, +∞)的float类型值

关联代码 #

Compute Shader 完整逻辑 #

[numthreads(X_THREADS, Y_THREADS, Z_THREADS)]
void CSMain (uint3 id : SV_DispatchThreadID)
{
    uint tid = id.x;
    while (tid < octree_length)
    {
        uint depth = OctreeDepth(tid);
        uint mid = tid - depth_offsets[depth];

        float node_size = bounds_size / pow(2, depth);
        float3 node_min = MortonToXYZ(mid) * node_size; 

        octree[tid].anchor = node_min;
        octree[tid].size = node_size;

        tid += total_num_threads;
    }
}

C#端着色器调度与结果回读代码 #

ComputeBuffer BuildOctree(Vector3 boundsMin, float boundsSize, int octreeDepth, Vector3Int numThreadGroups) 
{
    int totalNumThreads = numThreadGroups.x * numThreadGroups.y * numThreadGroups.z * THREADS_PER_GROUPS;

    int[] depthOffsets = new int[octreeDepth];
    int length = 0;
    for (int i = 0; i < octreeDepth; i++)
    {
        depthOffsets[i] = length;
        length += (int) Mathf.Pow(8, i);
    }

    ComputeBuffer octree = new ComputeBuffer(length, 4 * sizeof(float), ComputeBufferType.Structured);
    ComputeBuffer depthOffsetsGPU = new ComputeBuffer(octreeDepth, sizeof(int), ComputeBufferType.Structured);

    depthOffsetsGPU.SetData(depthOffsets);
    octree.SetData(new OctreeNode[length]);

    OCTREE_BUILDER.SetBuffer(0, "octree", octree);
    OCTREE_BUILDER.SetBuffer(0, "depth_offsets", depthOffsetsGPU);
    OCTREE_BUILDER.SetInt("total_num_threads", totalNumThreads);
    OCTREE_BUILDER.SetInt("octree_length", length);
    OCTREE_BUILDER.SetFloat("bounds_size", boundsSize);

    OCTREE_BUILDER.Dispatch(0, numThreadGroups.x, numThreadGroups.y, numThreadGroups.z);

    OctreeNode[] output = new OctreeNode[length];
    octree.GetData(output);

    for (int i = 0; i < output.Length; i++)
        Debug.Log("cell[" + i + "]: " + output[i].anchor + ", " + output[i].size);

    return octree;
}

修复方案 #

问题根源是HLSL编译器做隐式类型转换时，错误地为uint3转浮点的中间值分配了低精度浮点格式（半精度/定点数），导致计算结果偏差。

1. 核心修复：乘法前手动将uint3返回值显式转换为32位全精度float3，强制硬件使用全精度完成运算：

   float3 node_min = (float3)MortonToXYZ(mid) * node_size;
   

2. 附加精度优化：原代码中pow(2, depth)是整数次2的幂运算，调用通用浮点幂函数会引入不必要的精度损失，替换为专用实现效率和精度更高：

   // 方案1：位运算计算缩放值
   float node_size = bounds_size * (1.0f / (1u << depth));
   // 方案2：使用标准库ldexp函数直接计算2的幂缩放
   float node_size = ldexp(bounds_size, -(int)depth);
   

内容的提问来源于stack exchange，提问作者niv shalom