我的 OpenCL 内核抛出了一个浮点异常。我已将其简化为我认为导致问题的行。
如果我更换线路
acc.x += sin(distSqr);
和
acc.x += cos(distSqr);
或者
acc.x += sqrt(distSqr);
要不就
acc.x += (distSqr);
内核运行良好。为什么?注意:我的全球工作规模可以被我的本地工作规模整除。
谢谢。
这是内核:
__kernel void compute_forces(
__global float3 *x3a,
__global float3 *p3a,
__global float3 *x3b,
__global float3 *p3b,
__global float3 *f3a,
float dt,
float qQa,
float qQb,
float qma,
float qmb,
int n0a,
int n1a,
int n0b,
int n1b,
float xmin,
float ymin,
float epsSqr,
float force_fac,
__local float3 *localx
)
{
//we are going to compute the force between parts (n0a-n1a) and (n0b-n1b)
//Each particle loads the particle in the current block into local memory, so
unsigned int tid = get_local_id(0);
unsigned int gid = get_global_id(0);
unsigned int ninter=0;
// position of this work-item
float3 myPos = x3a[gid];
float3 acc = (float3)(0.0f, 0.0f, 0.0f);
// Synchronize to make sure data is available for processing
barrier(CLK_LOCAL_MEM_FENCE);
for(int j = 0; j < 2; ++j)
{
float3 r=-myPos;
float distSqr = r.x * r.x;
// accumulate effect of all particles
acc.x += sin(distSqr);
ninter++;
}//j
// Synchronize so that next tile can be loaded
barrier(CLK_LOCAL_MEM_FENCE);
f3a[gid]+=acc;
f3a[gid].x=(float)ninter;
}
我称内核为:
err=clSetKernelArg(k_compute_forces, 0, sizeof(_x3), &_x3);
err=clSetKernelArg(k_compute_forces, 1, sizeof(_p3), &_p3);
err=clSetKernelArg(k_compute_forces, 2, sizeof(_x3), &_x3);
err=clSetKernelArg(k_compute_forces, 3, sizeof(_p3), &_p3);
err=clSetKernelArg(k_compute_forces, 4, sizeof(_f3), &_f3);
err=clSetKernelArg(k_compute_forces, 5, sizeof(dt_float), &dt_float);
err=clSetKernelArg(k_compute_forces, 6, sizeof(qQa), &qQa);
err=clSetKernelArg(k_compute_forces, 7, sizeof(qQb), &qQb);
err=clSetKernelArg(k_compute_forces, 8, sizeof(qma), &qma);
err=clSetKernelArg(k_compute_forces, 9, sizeof(qmb), &qmb);
err=clSetKernelArg(k_compute_forces,10, sizeof(n0a), &n0a);
err=clSetKernelArg(k_compute_forces,11, sizeof(n1a), &n1a);
err=clSetKernelArg(k_compute_forces,12, sizeof(n0b), &n0b);
err=clSetKernelArg(k_compute_forces,13, sizeof(n1b), &n1b);
err=clSetKernelArg(k_compute_forces,14, sizeof(xmin_float), &xmin_float);
err=clSetKernelArg(k_compute_forces,15, sizeof(ymin_float), &ymin_float);
err=clSetKernelArg(k_compute_forces,16, sizeof(epsSqr), &epsSqr);
err=clSetKernelArg(k_compute_forces,17, sizeof(force_fac), &force_fac);
err=clSetKernelArg(k_compute_forces,18, parts_per_block*sizeof(cl_float3),NULL);
err=clEnqueueNDRangeKernel(queue, k_compute_forces, work_dim, NULL, global_work_size, local_work_size, 0, NULL, &k_compute_forces_completion);
编辑:我认为 sin 函数不能处理小于约 1.0e-12 的浮点数,因为该行:
acc.x += sin(1.0e-12);
运行良好,但
acc.x += sin(1.0e-13);
引发异常。这似乎表明正在调用 sin_half 而不是 sin...我想知道这是否是优化的替代品。
如果我在上述行之前添加一个 printf 语句
printf("distSqr=%g\n",distSqr);
然后错误从“浮点异常”变为有关“divisionErrorHandler”的内容(尽管由于输出文本混乱而难以辨认)。