Commit f74c9f43d934211679337c448361179e53a28bfe
1 parent
a3776b35
linux bugs involving nvcc for GPU-based hsiproc
Showing
1 changed file
with
2 additions
and
3 deletions
Show diff stats
stim/envi/convert.cu
1 | -#include <iostream> | ||
2 | -#include <stim/cuda/cudatools/error.h> | 1 | +//#include <stim/cuda/cudatools/error.h> |
3 | 2 | ||
4 | __global__ void kernel_permute(char* dest, char* src, size_t sx, size_t sy, size_t sz, size_t d0, size_t d1, size_t d2, size_t typesize){ | 3 | __global__ void kernel_permute(char* dest, char* src, size_t sx, size_t sy, size_t sz, size_t d0, size_t d1, size_t d2, size_t typesize){ |
5 | size_t xi = blockIdx.x * blockDim.x + threadIdx.x; | 4 | size_t xi = blockIdx.x * blockDim.x + threadIdx.x; |
@@ -25,4 +24,4 @@ void gpu_permute(char* dest, char* src, size_t sx, size_t sy, size_t sz, size_t | @@ -25,4 +24,4 @@ void gpu_permute(char* dest, char* src, size_t sx, size_t sy, size_t sz, size_t | ||
25 | dim3 threads(sqrt(threads_per_block), sqrt(threads_per_block), 1); | 24 | dim3 threads(sqrt(threads_per_block), sqrt(threads_per_block), 1); |
26 | dim3 blocks(sx/threads.x + 1, sy/threads.y + 1, sz/threads.z +1); | 25 | dim3 blocks(sx/threads.x + 1, sy/threads.y + 1, sz/threads.z +1); |
27 | kernel_permute<<<blocks, threads>>>(dest, src, sx, sy, sz, d0, d1, d2, typesize); | 26 | kernel_permute<<<blocks, threads>>>(dest, src, sx, sy, sz, d0, d1, d2, typesize); |
28 | -} | ||
29 | \ No newline at end of file | 27 | \ No newline at end of file |
28 | +} |